SmartRide: Intelligent reservation and scheduling for elevators

Abstract

With the continuous growth of the global population and the concentration of people in urban areas, the development of high-rise buildings has thrived, making elevator waiting time a significant concern. However, traditional elevators are unable to know passengers’ destination floors and waiting times in advance. Consequently, passengers often experience longer waiting times. With the flourishing advancement of the Internet of Things (IoT) and edge computing, it has become feasible to know the demands of passengers in advance and how long they have been waiting. In this work, we propose SmartRide, an intelligent elevator system that provides reservation, check-in, boarding, and scheduling features. SmartRide allows users to reserve elevators in advance through a mobile app from anywhere. Artificial intelligence (AI) image recognition technology is employed for efficient check-in and boarding confirmation. Additionally, a novel scheduling algorithm is proposed to minimize the average waiting time by leveraging the reservation information. Experimental results demonstrate that SmartRide can accurately manage check-ins and boarding confirmations. SmartRide can also reduce waiting times by at least 30% compared to traditional elevator scheduling using FCFS (First Come First Served), thereby optimizing elevator operational efficiency and improving user experience.

Keywords

elevators artificial intelligence facial recognition scheduling

1. Introduction

1.1. Current situation and problem

In recent years, as the global urbanization process continues to advance, the issue of elevator waiting has gradually emerged as a growing challenge. Elevators, as a crucial component of vertical transportation, will play an indispensable role in future cities. According to Al-Kodmany’s study on high-rise buildings [1], by the year 2030, 60% of the global population will reside in urban areas, possibly reaching 80% by 2050. With the ongoing increase in population density, cities can only accommodate more people by raising the height of buildings. According to statistics from the Council on Tall Buildings and Urban Habitat, there are already over 2,000 skyscrapers worldwide with heights exceeding 150 meters, and the number of buildings over 100 meters is even greater. As urbanization intensifies, high-density populations will concentrate within limited spaces, making transportation a highly challenging issue. According to market analysis by Grand View Research, the elevator market reached 19.8 billion USD in 2021 and is projected to reach 26.5 billion USD by 2025 [2]. Within the field of elevator research, issues are typically categorized into hardware and software aspects. Hardware aspects focus primarily on optimizing the elevator’s physical structure to enhance operational efficiency, energy-saving technologies, and safety improvements, including emergency systems for power outages and automated rescue devices. On the other hand, software aspects primarily involve the optimization of smart scheduling algorithms and user experience, such as the application of software interfaces, apps, or other technologies to enhance passengers’ overall experience. Our work falls under the software category, with a focus on exploring solutions to reduce waiting times and optimize the overall efficiency of elevator operations. Ongoing elevator-related research studies continuously strive to optimize elevator performance, with a dedicated focus on delivering an enhanced user experience. Among the issues addressed in elevator research, waiting times are the factor most directly influencing user experience. Prolonged waiting times for elevator services have consistently been linked to decreased customer or user satisfaction in numerous research papers [3 –7]. In addition, with the increasing height of buildings, elevators face unprecedented challenges, potentially resulting in significantly longer waiting and travel times for passengers. This not only affects residents’ quality of life but could also have substantial impacts on urban operations. Consequently, effectively addressing elevator waiting issues becomes a critical aspect for high-rise buildings.

Some studies have pointed out the elevator waiting problem that is often encountered in various fields. A previous study conducted by professional elevator consultant SVM Associates and corporate health business StepJockey showed that in commercial buildings in the UK, employees spend an average of at least 15 minutes a day waiting for elevators [8], which shows that optimizing passenger waiting time is very important. In addition, in a specific field, the problem of elevator waiting may also affect safety. Taking the related research of hotels as another example, some occasions may need to rely on elevators for evacuation, making the problem of waiting for elevators particularly critical [9]. In this situation, ensuring the efficient operation of the elevator system is not only related to the comfort experience of the residents, but also directly related to the safety of people’s lives. In the context of related research in hospitals, the operation of elevators becomes particularly crucial. Consider the implications of a situation where a doctor urgently needs to travel to a different floor for surgery or emergency treatment, and there is no elevator available at that moment. If the waiting time for the elevator is excessively long, the doctor may not be able to arrive in time [10], consequently delaying the patient’s treatment and potentially jeopardizing their life. Therefore, addressing these challenges necessitates the development of smarter elevators aimed at reducing waiting times.

1.2. Existing solutions of intelligent elevators

Several methods for reducing elevator waiting times have been proposed. Prior to the widespread adoption of IoT technology, elevator scheduling optimization primarily relied on traditional algorithms. Nail, A et al. [11] mentioned in their research that elevator algorithms have evolved from first-come-first-served (FCFS) to shortest floor search time first (SSTF), and then to scanning algorithms such as SCAN and LOOK. To enhance elevator scheduling efficiency, recent research has focused on the development of intelligent control algorithms, especially those involving the coordination of two or multiple elevators to address peak and off-peak crowd conditions. These methodologies include fuzzy control [12 –14], genetic algorithms [15 –18], neural networks [19], and reinforcement learning [20,21]. These algorithms establish control rules or models based on historical elevator data to predict floor requests during specific time periods, thus facilitating effective elevator scheduling. Consequently, these intelligent control systems require corresponding hardware equipment and substantial model training and testing efforts.

With the advancement of the Internet of Things, research has started to integrate sensor devices with elevator scheduling, introducing a range of possibilities for elevator optimization. For instance, Strang et al. [22] were the first to propose the concept of using Radio-frequency identification (RFID) devices to acquire destination floor information. Masaaki Amano et al. [23] redesigned the elevator’s external request panel, replacing directional arrows with buttons for all floors, enabling the system to proactively identify user starting and destination floors for more intelligent scheduling. G. Hangli et al. [24] expanded the information available on the request panel, allowing passengers to view the elevator’s current usage status and queue length, empowering users to decide whether to opt for the stairs. Wenxing Chen et al. [25] took a step further by integrating facial recognition functionality into the elevator request panel to verify the identity of reserved passengers through identification. Additionally, Khonjun et al. [26] proposed an elevator reservation system architecture where users can send reservation requests to a server using mobile devices, and the server optimizes scheduling for energy efficiency while controlling elevator movement sequences. In summary, traditional elevator scheduling algorithms relied on historical data to predict demand. However, with the advancement of IoT and smartphone technology, it has become possible to obtain user demand in advance. There is a gap in prior research on how to integrate users’ demands with elevator scheduling to further reduce waiting time and improve user experience.

1.3. Our contributions

In this paper, we propose a novel smart elevator system called SmartRide, which includes features of reservation, check-in, boarding, and scheduling. The prototype system is built on the College of Engineering Building 5 at National Yunlin University of Science and Technology. Its primary objective is to offer elevator reservation services to users while integrating a LOOK+ scheduler to reduce elevator waiting times. The SmartRide system employs IoT devices and incorporates a comprehensive user recognition mechanism to handle various steps in the user reservation process, including reserve, check-in, and board, for acquiring user-related information. LOOK+ algorithm then utilizes the information provided by users during the reservation process to schedule elevator operations. SmartRide is particularly applicable in hotels, office buildings, and hospitals. The major contributions of this work are described as follows.

(1) SmartRide offers a personnel recognition method using Jetson Nano and Raspberry Pi Model B3 development boards combined with a webcamera. The YOLOv7 image recognition model is integrated into the development boards to perform personnel recognition. Cameras are installed in corridors, lobbies, and inside elevators. The personnel recognition process involves three steps. Firstly, in the reserve phase, users can make elevator reservations through a designed app. The system acquires user-requested floor information, including user facial data, initial floor, and destination floor. Secondly, during the check-in phase, when users reach the lobby areas, such as corridors, lobbies, and elevators, cameras near the elevators confirm user identities. If users are waiting in the lobby area, the system records their waiting time for elevator scheduling purposes. Lastly, during the boarding phase, once users enter the elevator as guided by the app, SmartRide detects users entering and exiting the elevator through the camera inside the elevator, verifying the number of occupants and user identities.

(2) For the elevator scheduling algorithm, SmartRide proposes the LOOK+ algorithm, which is improved based on the most common LOOK algorithm for elevators, and is mainly used to prioritize passenger waiting time. The system will conduct scheduling by analyzing factors such as the passenger’s reserved floor, the current location of the elevator, the direction of the elevator, the capacity of the elevator, and the waiting time to reduce the average passenger waiting time. In addition, LOOK+ also combines the concept of Grouping, which combines the tasks of the same starting floor and destination floor, reducing the number of times the elevator moves up and down. Once the system determines the most suitable elevator, it will immediately assign this elevator to the passenger’s floor, allowing passengers to enter the elevator in a shorter time.

(3) Regarding implementation and field verification, this research conducted two parts of experiments. The first is the accuracy of personnel identification. In order to verify that SmartRide can provide effective information to the algorithm, the experiment simulated the identification results when the reservation person went to the elevator. The experiment proved that the accuracy of personnel identification was as high as 86%. During the process, passengers can be accurately identified and the correct information on the schedule can be provided. In the other part, we conducted three elevator simulation experiments for the LOOK+ algorithm: single elevator, multiple elevators, and different floor heights. Experimental results show that compared to traditional elevator systems, the LOOK+ algorithm’s elevator scheduling that takes waiting time into consideration can reduce user waiting time by at least 30%. As the number of floors increases, the reduction effect of waiting time is more significant.

The Sections of this paper are organized as follows. Section 2 reviews research on elevator scheduling and sensor applications. Section 3 discusses the design requirements and challenges of SmartRide. In Section 4, the system design and implementation of SmartRide are described. Section 5 analyzes the experiments on the accuracy of personnel identification and the effectiveness of the algorithm in SmartRide. Finally, Section 6 concludes this study.

2. Related work

There have been numerous studies dedicated to optimizing elevator control. The research in the field of elevator control can be broadly categorized into two main themes: mathematical optimization and sensor-based approaches. This section discusses elevator scheduling through mathematical optimization and IoT-enabled smart elevator applications, as described below.

2.1. Elevator scheduling

Table 1
A comparative table of elevator scheduling methods.

<

Author/year Methods Dynamic scheduling Optimizing passenger average waiting time Optimizing elevator total travel time Implementation of reservation (advance information acquisition)

Nail, A et al. / 2017 [11] FCFS, SSTF, SCAN/LOOK × × × ×

R. Gudwin et al. / 1998 [12] Fuzzy control ✓ ✓ ✓ ×

Chang Bum Kim et al. / 1998 [13] Fuzzy control ✓ ✓ ✓ ×

J. Fernández et al. / 2013 [14] Fuzzy control ✓ ✓ × ×

A. Fujino et al. / 1997 [15] Genetic algorithm ✓ ✓ × ×

Pablo Cortés et al. / 2004 [16] Genetic algorithm ✓ ✓ × ×

Z. Hu et al. / 2010 [17] Genetic algorithm ✓ ✓ × ×

E. O. Tartan et al. / 2014 [18] Genetic algorithm ✓ ✓ ✓ ×

İmrak, C. et al. / 2006 [19]< Neural network ✓ ✓ ✓ ×

R. H. Crites et al. / 1996 [20] Reinforcement learning ✓ ✓ ✓ ×

M. Ikuta et al. / 2013 [21] Reinforcement learning ✓ ✓ ✓ ×

G. Hangli et al. / 2020 [24] Reinforcement learning ✓ ✓ ✓ ×

Lan-Da Van et al. / 2020 [27] ACSICD ✓ ✓ ✓ ×

SmartRide / 2023 Grouping LOOK+ ✓ ✓ ✓ ✓

Author/year	Methods	Dynamic scheduling	Optimizing passenger average waiting time	Optimizing elevator total travel time	Implementation of reservation (advance information acquisition)
Nail, A et al. / 2017 [11]	FCFS, SSTF, SCAN/LOOK	×	×	×	×
R. Gudwin et al. / 1998 [12]	Fuzzy control	✓	✓	✓	×
Chang Bum Kim et al. / 1998 [13]	Fuzzy control	✓	✓	✓	×
J. Fernández et al. / 2013 [14]	Fuzzy control	✓	✓	×	×
A. Fujino et al. / 1997 [15]	Genetic algorithm	✓	✓	×	×
Pablo Cortés et al. / 2004 [16]	Genetic algorithm	✓	✓	×	×
Z. Hu et al. / 2010 [17]	Genetic algorithm	✓	✓	×	×
E. O. Tartan et al. / 2014 [18]	Genetic algorithm	✓	✓	✓	×
İmrak, C. et al. / 2006 [19]<	Neural network	✓	✓	✓	×
R. H. Crites et al. / 1996 [20]	Reinforcement learning	✓	✓	✓	×
M. Ikuta et al. / 2013 [21]	Reinforcement learning	✓	✓	✓	×
G. Hangli et al. / 2020 [24]	Reinforcement learning	✓	✓	✓	×
Lan-Da Van et al. / 2020 [27]	ACSICD	✓	✓	✓	×
SmartRide / 2023	Grouping LOOK+	✓	✓	✓	✓

As Table 1 shows, numerous research endeavors have been dedicated to optimizing elevator scheduling using mathematical algorithms. The primary objective is to enhance the efficiency of elevator dispatching by refining existing algorithms and optimizing various pertinent metrics, such as Average Waiting Time (AWT), Average Travel Time, Average Journey Time, and energy consumption. The commonly utilized elevator scheduling algorithms comprise FCFS, SSTF, SCAN, and LOOK.

FCFS is a straightforward first-come-first-serve algorithm, renowned for its fairness and simplicity. Nonetheless, it lacks scheduling optimization, and its performance significantly deteriorates under high request loads. SSTF prioritizes the nearest floor for the subsequent movement, thereby reducing elevator travel distance. However, in scenarios with continuous high request rates, certain floors may experience substantial waiting times. SCAN involves continual back-and-forth movement between the lowest and highest floors. This approach minimizes the number of round trips and stabilizes waiting times for various floors. LOOK is an enhancement over SCAN, in which the elevator moves between the lowest and highest floors, but it changes direction only after processing requests in the same direction. These conventional algorithms are chiefly suitable for low-floor, low-traffic applications. LOOK, in particular, finds common use in modern settings. Nevertheless, for tall buildings with high traffic, unpredictable traffic patterns, and the requirement for dynamic scheduling adjustments, these traditional algorithms may prove insufficient.

In recent years, elevator algorithms have undergone continuous enhancements. For example, Cortés et al. [28] introduced artificial intelligence methods to elevator control. Initial approaches encompassed fuzzy control methods [12 –14], where fuzzy rule sets were formulated based on observations and recordings of elevator operations. Subsequently, the system could translate the current elevator load into corresponding operational modes to address peak and off-peak demands. Genetic algorithms [15 –18] employed mutations and evolutions to explore all conceivable elevator operational scenarios until an optimal solution, adhering to predefined conditions, was achieved.Reinforcement learning [20,21,23] aimed to interact with the dynamic environment and learn continuously, adjusting parameters based on rewards received for favorable or unfavorable performance. This process enabled elevators to anticipate peak and off-peak periods and enhance scheduling accordingly. Furthermore, other studies have proposed their own methods, such as the elevator scheduling algorithm by Lan-Da Van et al. [27]. This algorithm selects the most suitable elevator for handling passenger requests based on each elevator’s status, considering movement direction and cabin occupancy information. Simulation results exhibited promising performance for these intelligent algorithms.

However, in practical applications, establishing and fine-tuning the functions in these algorithms necessitate substantial expert experience. The training process involves numerous unknown variables and demands robust hardware to handle intensive training. Moreover, existing elevator algorithm research has yet to address booking functionality. This paper presents a system that utilizes smartphones for advanced elevator reservation, revolutionizing the traditional method of obtaining elevator information solely from the up and down buttons at the elevator entrance. This innovation eliminates the need to rely on limited information for elevator scheduling.

2.2. IoT-enabled smart elevator applications

On the other hand, some research has focused on enhancing elevator efficiency by incorporating sensor technologies to gather more passenger- or elevator-related information (please refer to Table 2). For instance, Markku Turunen et al. [29] proposed the concept of using smartphones to reserve elevators. Their paper presented innovative reservation interfaces, but it did not disclose the implementation details of the reservation system. Ohhoon Kwon et al. [30] suggested an elevator system utilizing RFID, cameras, and floor sensors. RFID was used to obtain passenger information, while cameras and floor sensors confirmed whether passengers intended to board the elevator. The system would then proactively schedule elevators based on passenger habits. However, this approach may be less practical due to potential prediction errors, leading to increased waiting times. Wenxing Chen et al. [24] presented a comprehensive elevator reservation system, which is most closely related to the current study. Their paper not only proposed their own algorithm but also integrated UWB positioning systems to facilitate the reservation process. Nevertheless, from the user’s perspective, the verification process could be somewhat cumbersome. Assuming multiple passengers queued up for elevator entry on a given day, each passenger would need to undergo facial verification at the elevator panel, potentially causing delays. Rojanee Homchalee et al. [31] introduced a remote-controlled elevator system and presented their algorithm. However, the study lacked a personnel verification mechanism. Ensuring the correspondence between the person entering the elevator and the reserved passenger is crucial; otherwise, the scheduling efforts would become meaningless.

Table 2
IoT-enabled smart elevator applications.

Author/year Purpose Hardware equipment Methods

M. Turunen et al. / 2013 [29] Reduce waiting time 1.Smart phone Wireless link between mobile devices and elevators enables app-based reservations.

O. Kwon et al. / 2014 [30] Reduce waiting time 1. RFID sensor 2. Camera 3. Floor sensors Advance booking with RFID, cameras, and floor sensors gathers passenger data.

G. Hangli et al. / 2020 [24] Reduce waiting time 1. Smart phone 2. Elevator panels UWB and algorithms guide users to the nearest elevator.

R. Homchalee et al. / 2023 [31] Reduce waiting time 1. Smart phone 2. Elevator panels 3. Camera 4. Infrared sensors Algorithms use cameras and infrared sensors to optimize elevator dispatch.

D. Andriana et al. / 2021 [32] Enhance safety 1. Camera In-cabin cameras with Yolov3 ensure passenger spacing.

S.-C. Lai et al. / 2022 [33] Enhance safety 1. Arduino development board 2. Non-contact sensing button 3. Gesture recognition sensor Diverse sensors enable contactless elevator control.

E. Rubies et al. / 2023 [34] Enhance safety 1. Smart phone 2. Esp32 development board Modified button panels allow QR code-based elevator summoning.

S.M. Aljadani et al. / 2022 [35] Enhance safety 1. Temperature sensor 2. Infrared sensor 3. Photoresistor 4. Ultrasonic sensor 5. Vibration sensor6. Weight sensor 7. Liquid crystal display (LCD) Real-time sensor data aids maintenance troubleshooting.

P. Chandirasekeran et al. / 2022 [36] Enhance safety 1.Camera Cameras detect obstacles, triggering alerts for passengers.

Y. Zhang et al. / 2018 [37] Improve comfortability 1. Smart phone Smartphone accelerometers assess ride comfort via vibrations.

T. Matsumoto et al. / 2020 [38] Provide spatial guidance 1. LED floor 2. CreateLED AirMAG-3 display 3. Camera 4. Hokuyo UST-10LX-H01 scanning lidar Displays and tracking guide passengers in elevator boarding.

Author/year	Purpose	Hardware equipment	Methods
M. Turunen et al. / 2013 [29]	Reduce waiting time	1.Smart phone	Wireless link between mobile devices and elevators enables app-based reservations.
O. Kwon et al. / 2014 [30]	Reduce waiting time	1. RFID sensor 2. Camera 3. Floor sensors	Advance booking with RFID, cameras, and floor sensors gathers passenger data.
G. Hangli et al. / 2020 [24]	Reduce waiting time	1. Smart phone 2. Elevator panels	UWB and algorithms guide users to the nearest elevator.
R. Homchalee et al. / 2023 [31]	Reduce waiting time	1. Smart phone 2. Elevator panels 3. Camera 4. Infrared sensors	Algorithms use cameras and infrared sensors to optimize elevator dispatch.
D. Andriana et al. / 2021 [32]	Enhance safety	1. Camera	In-cabin cameras with Yolov3 ensure passenger spacing.
S.-C. Lai et al. / 2022 [33]	Enhance safety	1. Arduino development board 2. Non-contact sensing button 3. Gesture recognition sensor	Diverse sensors enable contactless elevator control.
E. Rubies et al. / 2023 [34]	Enhance safety	1. Smart phone 2. Esp32 development board	Modified button panels allow QR code-based elevator summoning.
S.M. Aljadani et al. / 2022 [35]	Enhance safety	1. Temperature sensor 2. Infrared sensor 3. Photoresistor 4. Ultrasonic sensor 5. Vibration sensor6. Weight sensor 7. Liquid crystal display (LCD)	Real-time sensor data aids maintenance troubleshooting.
P. Chandirasekeran et al. / 2022 [36]	Enhance safety	1.Camera	Cameras detect obstacles, triggering alerts for passengers.
Y. Zhang et al. / 2018 [37]	Improve comfortability	1. Smart phone	Smartphone accelerometers assess ride comfort via vibrations.
T. Matsumoto et al. / 2020 [38]	Provide spatial guidance	1. LED floor 2. CreateLED AirMAG-3 display 3. Camera 4. Hokuyo UST-10LX-H01 scanning lidar	Displays and tracking guide passengers in elevator boarding.

Additionally, regarding safety aspects, Dian Andriana et al. [32] proposed real-time passenger distance detection, which could monitor passenger numbers and distances. Nevertheless, its application might be challenging beyond epidemic-related scenarios. Shu-Chen Lai and Elena Rubies et al. et al. [33,34] suggested a contactless elevator system that allows various touchless methods, offering significant benefits for epidemic prevention and convenience. However, in real-life situations, not everyone may opt for using a touchless system, as many individuals might still prefer directly pressing the buttons. Sarah et al. [35] proposed real-time monitoring of elevator operational status using multiple sensors and automatic reporting of malfunctions. Preethi Chandirasekeran et al. [36] suggested utilizing cameras to detect obstacles during elevator door closure and promptly sending alerts. Other applications of the Internet of Things (IoT) in elevators include Yang Zhang et al.’s [37] method of using smartphones’ built-in accelerometer to monitor ride comfort. Nevertheless, detecting comfort levels might not be very practical for general passengers. Takashi Matsumoto et al. [38] introduced the use of large floor displays to guide passengers to the appropriate elevator, a novel approach that would require high-cost hardware to be realized.

Different from the existing algorithms and sensor deployment methods, SmartRide improves the LOOK algorithm in order to reduce the waiting time. The information required by the algorithm including identity, destination floor, waiting time and number of people in elevator, is obtained through the use of reservation APP and low-cost development board with camera combined with YOLOv7 image recognition technology. SmartRide provides a more complete way to obtain information, and is also easy to configure and maintain.

3. Design requirements

The objective of this study is to design an intelligent elevator system with reservation functionality, providing users with an app to make elevator reservations. The system will employ algorithms for scheduling based on the information gathered from user reservations, aiming to reduce waiting times. Additionally, it is crucial to verify the arrival status of users after making a reservation to ensure the system’s integrity. The following sections will describe the algorithm design, personnel identification mechanism, and app features separately.

3.1. Minimizing waiting time

In algorithm design, scheduling needs to consider the priority order of all passenger waiting times. It is important to avoid a scenario where the elevator only prioritizes the shortest travel distance or shortest travel time when handling all requests, while neglecting the fact that a request from a certain floor may have gone unattended for a considerable amount of time, resulting in an overall longer waiting time. Therefore, this study aims to strike a balance in the overall waiting time, allowing the scheduling to address all requests and reach their destinations while minimizing the waiting time for each passenger as much as possible.

3.2. Providing reservation and auto recognition

To implement the proposed elevator reservation feature discussed in this paper and improve the overall user experience and user-friendliness of the operation, we have outlined the requirements for this reservation feature as follows. To utilize the reservation feature effectively, users must log in or register within the reservation system and provide accurate floor request information before they can proceed with making a reservation. Regarding the cancellation of reservations, there should be a mechanism in place to facilitate the cancellation process. Reservations can be canceled by user-initiated actions. This measure prevents scenarios in which passengers make reservations but do not show up, thereby avoiding disruptions to the scheduling. To recognize when users have arrived, an automated mechanism should be implemented after users have made their reservations. This mechanism will determine if users have reached the elevator area. For the starting point of pedestrian recognition, as pedestrians pass through the corridor leading to the elevator, the system needs to identify the appropriate point at which to begin pedestrian recognition. This step is crucial for confirming whether the pedestrian matches the identity of the reserved passenger. By incorporating these design requirements, our objective is to create a comprehensive reservation feature that not only enhances user experience and efficiency but also ensures the integrity of the algorithmic functionality.

4. Methodology

4.1. System architecture

Figure 1.

User context.

This paper introduces an intelligent elevator system, SmartRide, equipped with a reservation feature. In order to optimize elevator scheduling, this system assumes that all individuals must use the reservation method to ride the elevator and enter it according to the system’s instructions. Figure 1 illustrates the scenario of users reserving the elevator. Users need to go through a three-step process: reservation, check-in, and boarding. Initially, users can utilize the developed app anywhere within the building to scan their faces and make a reservation, providing algorithmic booking information. In the second step, prior to users approaching the elevator, the system utilizes a setup development board equipped with a camera to perform real-time detection of whether the user has entered the lobby area within the camera’s field of view (FOV). This process serves to confirm their identity and record the waiting time. Finally, in the third step, users enter the elevator, and the camera inside the elevator detects their presence, updating the count of individuals within the elevator. Similarly, the count is updated again when the user exits the elevator, completing the entire boarding process. Among these three applied scenarios, only the first step, reservation, requires manual user operation. The system automatically verifies the check-in and boarding in the second and third steps, respectively. Check-in confirmation marks the start of the waiting time calculation after users complete the check-in process. Boarding confirmation is intended to inform the algorithm that the task has been processed, eliminating the need for further scheduling. Therefore, all three steps provide the necessary information for the algorithm to update dynamic scheduling on each occasion and ensure its correctness.

Figure 2.

Development board.

The camera system of this project is composed of development boards and a standard 1920*1080 pixel webcamera. In this paper, we have chosen the Raspberry Pi 3 Model B and Jetson Nano Developer Kit B01 as the development boards, as shown in Figure 2. The specifications of these two development boards are detailed in Table 3. The main reason for selecting the Raspberry Pi 3 Model B and Jetson Nano Developer Kit B01 as the primary platforms is their excellent cost-performance ratio. Firstly, the Raspberry Pi 3 Model B boasts outstanding performance. It is a powerful and cost-effective development board equipped with a fast quad-core processor, 1 GB of memory, and various I/O interfaces. These features make it highly suitable as the main controller for the camera system. Additionally, the Raspberry Pi 3 Model B benefits from extensive community support and a wealth of available resources, making system development and expansion much easier. Secondly, the Jetson Nano Developer Kit B01 was chosen as another development board due to its favorable cost-performance ratio. Designed specifically for artificial intelligence and machine learning applications, the Jetson Nano Developer Kit B01 features NVIDIA’s GPU with 128 CUDA cores, enabling excellent performance in image processing and machine learning tasks. Not only does the Jetson Nano Developer Kit B01 offer powerful computing capabilities, but it also has low power consumption and a relatively affordable price. Hence, the Jetson Nano Developer Kit B01 stands as an ideal choice for developing camera systems based on image processing and machine learning. In conclusion, both the Raspberry Pi 3 Model B and Jetson Nano Developer Kit B01 provide sufficient computing power and flexibility while being relatively cost-effective and enjoying extensive community support and abundant resources. This makes system development and expansion easier and more cost-efficient.

Table 3

Development board specifications.

	Raspberry Pi 3B	Jetson Nano B01
Performance	5.3 GFLOPs	472 GFLOPs
GPU	128-core MaxwellTM GPU	Dual Core VideoCore IV Multimedia Co-Processor
CPU	Quad-core ARM Cortex-A53	Quad-core ARM Cortex-A57 MPCore
GPU memory	1 GB	4 GB
Power	5 W	10 W
(L×W×H)(mm)	85×56×17	100×80×29
Price (USD)	45$	149$

4.2. Personnel recognition mechanism

To ensure the integrity of the user’s complete usage process, this study employs an on-site authentication mechanism to assist the system in decision-making. The following subsections elaborate on three steps: reserve, check-in, and board.

Figure 3.

Interface design for the mobile application (APP).

4.2.1. Reservation process

In order to provide users with a smoother elevator reservation experience, this paper develops a mobile phone application for reservation users to make elevator reservations. The interface design of the mobile phone application is shown in Figure 3, from left to right are (A) main interface, (B) reservation page, (C) reservation success page, (D) current usage status page and (E) usage history page. The following is an introduction to each interface. (A) Main interface: Before making an appointment, users who use the appointment system for the first time need to perform facial recognition. This step is designed to ensure that the cameras can identify passengers before they enter the elevator to determine the user’s entry and exit. After the user completes the facial recognition registration, the user does not need to perform this action the next time he uses the system, and can directly go to the appointment page to make an appointment. (B) Reservation page: This page is the main reservation page. Provide the selection fields of the current floor and the floor to go to, and you can choose the starting floor and the destination floor to be reserved. (C) Reservation success screen: When the reservation is successful, the system will jump to the reservation success screen, and the user can reconfirm whether his reservation floor is correct. The button below provides the function of canceling the reservation. Users can use this function to cancel the reservation when they do not need to use the elevator temporarily or want to make changes, so as not to affect the operation of the elevator system. (D) Current usage status page: This page displays the order of reservation tasks. Passengers can line up according to their order on the screen. When the elevator assigned by the system arrives, the user will be prompted to enter. (E) Use record page: This page will display the user’s reservation record, and provide the busy period of elevator use for the user’s reference.

4.2.2. Check-in process

The check-in process is a crucial component of the system, where cameras built using development boards are installed at elevator lobby and corridor locations to detect real-time images. Once a user successfully makes a reservation and appears within the camera’s detection range, the system performs facial recognition on the current image and compares it with the pre-trained YOLOv7 facial recognition model. If the system successfully identifies that the person in the image is the same as the reservation holder, upon the user’s arrival at the lobby, the system initiates the calculation of their waiting time, providing input for the algorithmic scheduling. This process ensures both accurate identity verification and equitable treatment of every passenger. Furthermore, this check-in procedure effectively reduces the occurrence of false reports. This is because the system only starts calculating the waiting time after verifying the facial match across multiple cameras’ monitoring ranges, thus preventing instances where others might be mistakenly identified as reservation holders.

4.2.3. Boarding process

During the boarding process, this paper also utilizes a camera mounted inside the elevator, which is integrated into the development board, to capture real-time images of the interior. Through the object detection module in the development board, the system can instantly detect the number of passengers inside the elevator and update this information in the system. This update action ensures the correctness of the scheduling algorithm since it relies on an accurate understanding of the number of passengers inside the elevator.

By employing real-time detection with the object detection module, the system can continuously monitor changes in the number of passengers inside the elevator. This ensures that the scheduling algorithm makes precise decisions based on the latest data. This system design provides a better experience for passengers while ensuring the correct operation of the entire system. The combination of the camera and the object detection module enables real-time monitoring and updates of the number of passengers inside the elevator, allowing the scheduling algorithm to make optimal decisions based on accurate data, thereby providing a more efficient and safe boarding experience.

4.3. Personnel recognition method based on cloud computing

The system architecture of the cloud-based personnel recognition method proposed in this paper is illustrated in Figure 4. It is divided into three main parts from top to bottom: Reserve, Check-in, and Board, which are introduced separately.

Figure 4.

Architecture diagram of personnel recognition method in cloud computing.

In the Reserve phase, users are required to verify their identity by scanning their face through a reservation app. First-time users need to register in advance, as explained in Section 4.2.1, and input their starting and destination floors, providing the system with user request information. Whenever a user makes a reservation, the database stores the new reservation data. The elevator status segment below retrieves the current floor, moving direction, and number of occupants inside the elevator. This information is updated each time the elevator moves or someone enters or exits. The database stores and updates both the aforementioned reservation information and elevator status for scheduling purposes. After the database adds a new reservation task and verifies the user input, a message indicating successful reservation is sent back to the user.

Moving on to the Check-in phase, at the elevator entrance and in the corridor, cameras built on development boards provide real-time images. When a user’s reservation is successful and they appear within the camera’s detection range, the pre-trained YOLOv7 face recognition module is employed to match the current image with the reserved user’s image, initiating facial recognition. When the camera at the elevator hall detects that the individual matches the reservation, the development board records the user’s waiting time and sends it to the database for scheduling purposes. Once a user has completed the check-in process, the database contains user reservation details, elevator status, and waiting time. After obtaining information from the above three points using the LOOK+ algorithm, the server conveys instructions to control the elevator controller in the order determined by the schedule, assigning appropriate elevator movement sequences.

The Board phase, which is also facilitated by a camera built on a development board. The server employs an object detection module to detect the number of occupants inside the elevator in real time, updating the number in the database to ensure the correctness of the algorithm. It’s important to note that, due to experimental constraints, the elevator controller in the second phase has not been modified in the actual field. Therefore, this paper utilizes algorithmic simulation to address this aspect.

This study employed the YOLOv7 convolutional neural network for user recognition in cloud based personnel identification [39]. YOLOv7, the latest version published in July of last year by Dr. Chien-Yao Wang and his team, optimized the model architecture for real-time object recognition by focusing on two main directions: model reparameterization and dynamic label assignment strategy. These optimizations resulted in faster training speed and higher accuracy compared to previous YOLOv5 versions, with a 20% increase in training speed. Since its release in the year 2022, YOLOv7 has provided more than 20 pre-trained models, including six common variants: YOLOv7, YOLOv7-X, YOLOv7-W6, YOLOv7-E6, YOLOv7-D6, and YOLOv7-E6E. Each variant represents a different training depth, which necessitates hardware capable of meeting their performance requirements. Among these models, YOLOv7 and YOLOv7-X achieved the highest official accuracy and are commonly used in experiments. Therefore, this study used YOLOv7 and YOLOv7-X for experimentation and compared their accuracy on the training set.Four students were invited to participate in the experiment, providing approximately 50 facial photos from various angles for each individual, resulting in a total of about 210 images for the dataset. The participants were labeled as user1, user2, user3, and user4 using the LabelImg image annotation tool. For the training and testing parts, the dataset was divided into approximately 80% for training (167 images) and 20% for testing (43 images).

The training parameters for YOLOv7 and YOLOv7-X in this experiment were configured as follows. First, the image size was set to 640*640, representing the dimensions of the input images to the model. The YOLO model divides the input image into smaller grids and predicts the object’s position and category for each grid. Second, the batch size was set to 16, indicating the number of samples processed in each training iteration. A larger batch size allows more samples to be used for training, providing a more stable gradient estimate and aiding in converging to better local minima. Finally, the epoch was set to 50, representing the number of times the entire training dataset passes through the neural network during training. Since the dataset used in this experiment is not large, 50 epochs were chosen for the training process. The training results are shown in Figures 5 and 6. “True Positives” represent the number of samples where both the actual ground truth and the model’s prediction are positive, abbreviated as TP. “False Positives” represent the number of samples where the actual ground truth is negative, but the model predicts a positive outcome, abbreviated as FP. “True Negatives” represent the number of samples where both the actual ground truth and the model’s prediction are negative, abbreviated as TN. “False Negatives” represent the number of samples where the actual ground truth is positive, but the model predicts a negative outcome, abbreviated as FN.

Figure 5.

Custom YOLOv7 training results.

Figure 6.

Custom YOLOv7X training results.

In Figure 5 and 6, from left to right, we have Precision, which indicates the proportion of true positive samples among the predicted positive samples, and its formula is Precision = TP / (TP + FP); Recall, which represents the proportion of true positive samples among the actual positive samples, and its formula is Recall = TP / (TP + FN); IoU (Intersection over Union), which denotes the overlap ratio between the predicted object detection box and the actual object, and its formula is IoU = (Intersection(A, B)) / (union(A, B)); mAP at IoU threshold 0.5, denoted as mAP0.5, represents the mean Average Precision when using an IoU threshold of 0.5. The experimental results in Tables 4 and 5 show the number of test images (Image) and the number of recognized labels (Labels). For YOLOv7, the Precision (P) is 0.996, Recall (R) is 0.881, and mAP0.5 is 0.961, with mAP0.5:0.95 reaching 0.799. For YOLOv7X, the Precision (P) is 0.985, Recall (R) is 0.93, and mAP0.5 is 0.981, with mAP0.5:0.95 reaching 0.803. This indicates that both models exhibit excellent recognition rates, with YOLOv7X being slightly more accurate than YOLOv7.

Table 4

Custom YOLOv7 training results data.

Class	Images	Lables	P	R	mAP@.5	mAP@.5:.95
All	43	52	0.996	0.881	0.961	0.799

Table 5

Custom YOLOv7X training results data.

Class	Images	Lables	P	R	mAP@.5	mAP@.5:.95
All	43	52	0.985	0.93	0.981	0.803

In the personnel recognition method for cloud computing, this study utilizes the Raspberry Pi 3 Model B in combination with a low-cost USB webcam to transmit real-time images to the server. The images are then subjected to relevant analysis and processing using pre-trained models. Initially, WebSocket was employed for real-time image transmission but encountered issues such as low transmission efficiency and delays. To overcome these problems, a tool named Mjpg-streamer was adopted for image compression and transmission. Mjpg-streamer is an open-source software capable of compressing images into MJPEG format and transmitting them to the server in real-time.

Using mjpg-streamer offers several advantages, including the effective reduction of image data size through compression algorithms, thereby reducing the required bandwidth and transmission time. This proves beneficial when dealing with limited network resources for image transmission. Moreover, mjpg-streamer provides a simple and intuitive user interface, making it easy to configure and run on the Raspberry Pi. Users can easily adjust compression parameters to balance image quality and transmission efficiency. Additionally, mjpg-streamer supports multiple protocols, including HTTP and TCP, allowing images to be transmitted to the server in different ways for subsequent processing and analysis. Consequently, this study chooses to use mjpg-streamer as the image compression and transmission tool to enhance the efficiency and quality of image transmission, thereby accelerating the overall system’s responsiveness and accuracy.

During the research process, it was discovered that although image compression and transmission optimization techniques were considered to reduce the amount of image data that the server needs to process, thus improving system performance and response speed, a crucial issue arose. The system needs to handle image data from multiple floors simultaneously. In practical elevator systems, multiple floors are involved, and each floor requires cameras to transmit real-time images back and forth. This implies that the system needs to process image data simultaneously from multiple development boards, and this image data can be substantial. This poses significant challenges to the server’s image processing capacity and network transmission speed and may lead to overload issues. To address this problem, it was necessary to reevaluate the system architecture and data processing strategy, delegating some image processing tasks to the development boards to assist in sharing the server workload.

Figure 7.

Software architecture diagram for personnel identification method in embedded AI.

4.4. Personnel recognition method based on embedded AI

In order to address the aforementioned challenge, this study replaces the Raspberry Pi 3 Model B with Nvidia Jetson Nano B01, as shown in Figure 7. The image recognition, previously processed on the server, is now performed on the development board. The server-side only receives the recognized information for updating the database.

Nvidia Jetson Nano B01 offers significant advantages over Raspberry Pi in terms of processing power, deep learning acceleration, software support, and peripheral expandability. Some advantages of Nvidia Jetson Nano B01 compared to Raspberry Pi are as follows. Processing Power: Nvidia Jetson Nano B01 is equipped with Nvidia’s GPU, providing powerful graphics processing and computational capabilities. In comparison, Raspberry Pi’s processing power is relatively weaker and mainly relies on its ARM-based CPU. AI Focus: Jetson Nano B01 is designed specifically for AI and machine learning applications. It comes with various software and tools optimized for AI applications, including TensorRT, cuDNN, CUDA, etc., making it more advantageous for AI-related tasks. Software Ecosystem Support: Nvidia Jetson Nano B01 supports Nvidia’s JetPack software package, which includes a wide range of development tools and libraries, facilitating the development of deep learning, image processing, and computer vision applications. In general, compared to the Raspberry Pi 3 Model B’s GPU, which is only suitable for general computing, the Nvidia Jetson Nano B01’s GPU has multiple CUDA cores, which allows it to handle multiple tasks at the same time. In image recognition, we allocated different image areas to different cores of the GPU for processing to achieve data parallelism. This parallel processing method helps to speed up image recognition, especially when batch processing of large amounts of image data, it can more effectively utilize the parallel computing power of the GPU.

In the personnel recognition method for embedded AI, this study utilizes YOLOv7-Tiny as the image recognition model for the embedded development board. Tiny models are known for their smaller sizes and fewer parameters, leading to lightweight object detection. This reduction in Tiny models is typically achieved through various methods. Tiny models simplify the network architecture, reducing the number of layers and feature extractor depth. This in turn reduces the model’s computation and parameter count. Additionally, spatial downsampling is reduced in Tiny models, preserving more local details and features. While this may slightly impact the detection of small objects, it remains acceptable for objects of general size. Moreover, Tiny models may adopt a smaller grid size, which divides the input image into fewer grid cells. This reduces the predicted positions of object centers but can also influence the detection of small objects. Despite the potential slight decrease in accuracy, Tiny models are highly favored for resource-constrained embedded devices and edge computing environments due to their lightweight nature. Consequently, this paper has selected YOLOv7-Tiny for implementing personnel recognition with embedded devices.

The embedded AI architecture uses the same dataset and training parameters as described in Section 4.3, setting the image size to 640*640, batch size to 16, and epochs to 50. The training results are shown in Figure 8, and Table 6 indicates an mAP@0.5 of 0.486. As it is a smaller model, the training speed is only about one-third of a regular model, resulting in accuracy that is relatively lower.

Figure 8.

Custom tiny YOLOv7 training results.

Table 6

Custom tiny YOLOv7 training results data.

Class	Images	Lables	P	R	mAP@.5	mAP@.5:.95
All	43	52	0.377	0.86	0.643	0.486

Figure 9.

YOLOv7 model training process.

To further enhance the model’s performance, it underwent multiple rounds of retraining and parameter adjustments during the training process, as depicted in Figure 9 below. The model training incorporated a precision-enhancing feature, specifically the multi-scale training option. This multi-scale training approach improved the model’s detection capabilities for targets of different sizes. In the real world, the dimensions of objects can vary significantly. Therefore, employing multi-scale training helps the model adapt better to a diverse range of target sizes. By randomly selecting various image sizes during training, the model learns a broader set of features associated with different scales. Furthermore, to better handle small-scale targets with the compact model, the experiment shifted from using 640x640 image dimensions to 1280x1280 dimensions. Subsequently, an evaluation was performed based on the results obtained from multi-scale training. As indicated in Table 7 and Figure 10, a significant increase in mean average precision was observed, demonstrating the effectiveness of these adjustments.

Table 7

Custom tiny YOLOv7 with multi scale training results data.

Class	Images	Lables	P	R	mAP@.5	mAP@.5:.95
All	43	52	0.96	0.985	0.993	0.841

Figure 10.

Custom tiny YOLOv7 with multi scale training results.

4.5. LOOK+ elevator algorithm

This paper proposes a novel elevator dispatching algorithm called LOOK+. The algorithm’s objective is to minimize elevator waiting time and enhance passenger satisfaction. The following presents a detailed description of the algorithm’s process. When a new task needs to be assigned, the system invokes the get_Request function, which returns the task’s $R_{i}$ ( $i = 1, 2, 3 \dots N$ ), starting floor $S_{i}$ , destination floor $D_{i}$ , and waiting time ${W T}_{i}$ (line 2). Subsequently, the system adds the task Request_i to the task queue (line 3) and sorts it in descending order of waiting time with Grouping procedure (line 4), as illustrated in Algorithm 1. The Grouping algorithm (line 20–34) will first classify all obtained requests (line 22–28), cluster the same $S_{i}$ , $D_{i}$ tasks into one task, and then calculate the waiting time of the task (line 29–33), and finally sort the priority. In this work, the total waiting time for each group of passengers is calculated by summing up the waiting times of individual passengers within the group. This cumulative waiting time is then divided by the total number of passengers in the group to obtain the average waiting time. We use the average waiting time of passengers in the same group for subsequent sorting.

An illustrate example is shown in Figure 11. For instance, if there is an empty elevator moving upwards with a capacity of 10 people, after sorting the waiting time of the tasks, the tasks with red numbers 1 to 10 with higher waiting times will be selected first. Then assign to the corresponding elevator.

Figure 11.

Diagram of priority-based processing.

For elevator allocation, the algorithm aims to enhance scheduling efficiency by prioritizing the assignment of tasks with longer waiting times to currently idle elevators. If there are more than one idle elevator, the system randomly selects one for task assignment (lines 5–7). In cases of frequent elevator usage, when all elevators are engaged and not in idle mode, the algorithm makes determinations (line 8) (line 19). Suppose an elevator is moving in the direction of a task’s pickup floor and the destination floor is in the same direction. In that case, the system prioritizes assigning the task to that elevator. Additionally, the system further evaluates based on the number of passengers and their destination floors within the elevator. As the elevator already has passengers on board, three scenarios might arise during the dynamic allocation of pending tasks. Firstly, if there are more vacant spaces in the elevator than required for the task (line 10), the system places the task within that elevator, considering Grouping (as shown in Figure 12), where passengers with the same starting and destination floors are treated as a single task, which might include more than one person. Secondly, if the vacant spaces in the elevator are fewer than required for the task (line 12), the system tries to accommodate a portion of the task’s passengers in the elevator to enhance utilization. Lastly, if the vacant spaces in the elevator exactly match the task’s requirements (line 15), the system directly assigns the task to that elevator. If an elevator is already at full capacity (line 16), the system shifts its consideration to the next elevator and repeats the task assignment decision process until all tasks are allocated to the most suitable elevators. To ensure the smooth operation of the algorithm for optimal elevator scheduling, maintaining an up-to-date and accurate database is essential. Hence, a robust on-site verification mechanism is necessary to guarantee precise and complete data, allowing the algorithm to function correctly and make optimal elevator scheduling decisions.

4.6. Discussion

In this study, apart from reducing waiting times, security management emerges as a paramount concern. To ensure user safety and the proper operation of the system, SmartRide employs a reservation system, obligating users to proceed through a reservation process before utilizing the system. Individuals who are unregistered or do not possess the requisite mobile device will be unable to avail themselves of SmartRide’s services. This measure guarantees that only authorized users are able to access and utilize the system, effectively eliminating the possibility of unauthorized usage. Furthermore, within specific application contexts, such as commercial establishments with high foot traffic and limited floors, we acknowledge that SmartRide may not be universally suitable.

However, practical implementation requires careful consideration of potential issues that might arise in specific scenarios. To address these concerns, we have devised strategies tailored to different situations. In the first scenario, where the risk of unauthorized access by tailgating behind a legitimate reservation holder is present, we have implemented preventive measures. Our solution involves equipping the system with real-time camera recognition technology to identify non-reservation holders. If a non-reservation holder is detected entering the elevator, the system will withhold elevator operation, keeping the doors open until the unauthorized individual exits. This effectively eliminates the possibility of queue jumping, ensuring user safety and a seamless experience. In the second scenario, we have also taken into consideration users who might not possess mobile devices or have difficulty operating them, such as elderly individuals. For such cases, we have contemplated the potential need for elevator attendants, like building security personnel, to provide assistance. These attendants can offer the necessary help, ensuring that users with special requirements can use the SmartRide system safely and effortlessly. In anticipation of exceptional situations, practical deployment might involve reserving one elevator exclusively for non-reservation holders, if the elevator capacity allows for it. In conclusion, the design of the SmartRide system comprehensively addresses security concerns. By incorporating reservation systems and personnel recognition mechanisms, it provide users with a secure and reliable operational environment, ensuring the smooth functioning of the system.

Figure 12.

Task grouping handling.

5. Experiments

5.1. Experiment design

To implement and validate SmartRide, the performance of personnel identification was evaluated under cloud computing architecture in Section 5.2, and under edge computing architecture in Section 5.3. A performance comparison between the two architectures was conducted in Section 5.4. Furthermore, algorithmic performance, presented in Section 5.5 was verified through simulations. Traditional elevator scheduling algorithms including FCFS, SSTF, and LOOK algorithms were selected for comparison with the proposed LOOK+ and LOOK+ combined with Grouping algorithms in this study. The evaluation criterion primarily focused on AWT, which represents the average time passengers wait from calling the elevator to its arrival. A shorter AWT indicates reduced waiting times for passengers, signifying higher elevator system efficiency. Therefore, AWT serves as a critical indicator for assessing elevator system performance. The experimental environment variables encompassed the number of floors, elevator count, elevator capacity, and number of passengers. Each passenger’s starting floor, destination floor, and waiting time were generated randomly and stored in arrays.

5.2. Personnel recognition accuracy and inference time based on cloud computing

In the context of cloud computing architecture, this experiment is focused on the validation of personnel recognition accuracy. To achieve this objective, real-time images were captured from strategic areas frequently visited by passengers. As illustrated in Figure 13, which depicts the experimental field setup, the diagram is a floor plan of the area near the elevators on the second floor of the Engineering Building 5 at National Yunlin University of Science and Technology. Areas A and B represent locations before and at the corner of the corridor, respectively. Area C corresponds to the lobby’s location, while area D indicates the position of the elevator. The Raspberry Pi cameras positioned in the elevator, lobby, and corridor were strategically utilized to capture these images. Through the utilization of a YOLOv7 trained model, the model will automatically recognize objects or features in the image and provide prediction results. Each boxed facial instance is concomitant with its associated user in the top-left corner, distinguished by distinct colors.

Nevertheless, due to various intervening factors such as lighting, angles, and attire, the model may encounter recognition inaccuracies under specific circumstances. For instance, it might struggle to demarcate individuals or inaccurately label them. To address such occurrences, manual inspection is conventionally employed. The precision of manual inspection necessitates additional validation of the model’s recognition results through human intervention, thus affirming the model’s accuracy and dependability. As illustrated in Table 8, the experiment involves the manual scrutiny of predicted videos. Each video encompasses a rate of 30 frames per second, with the following durations: (A)(B)Corridor, 29.7 seconds, roughly 891 frames; (C) Lobby, 14.8 seconds, roughly 444 frames; (D) Elevator interior, 10.1 seconds, roughly 303 frames. A comparative analysis between Predicted Results (PR) and Visual Inspection (VI) is executed, and accuracy is computed utilizing the equation Accuracy = (Number of PR = VI) / total frames. Table 8 encompasses four discrete scenario tests.

Figure 13.

Experimental field layout diagram: (A)(B) corridor (C) lobby (D) inside the elevator.

Table 8

Personnel identification accuracy based on cloud computing.

Experiment no.	Region no.	Test scenario	User1	User2	User3	User4
1	A	Corridor	68%	71%	56%	64%
2	B	Corridor-corner	83%	100%	81%	74%
3	C	Lobby	100%	100%	75%	77%
4	D	Inside the elevator	100%	89%	95%	100%

The insights drawn from the experimental outcomes in Table 8 and Figure 14 unveil two pivotal observations. Initially, recognition rates before the corner of the corridor are relatively modest. This can be attributed to the reality that faces at distant points in the corridor are frequently too diminutive and indistinct for efficacious recognition. Thus, for pragmatic applications, appropriate adjustments in camera positioning to ensure ample illumination and clarity are imperative to augment recognition precision. Additionally, a substantial determinant impacting precision pertains to the blurring of facial attributes during human motion. The motion of individuals can induce blurring or lack of focus in facial features, thereby rendering accurate detection and recognition arduous for facial recognition systems. This blurring is particularly conspicuous during swift or abrupt movements. For instance, in the scenario depicted in Figure 15 below, wherein user1 enters the elevator with lengthy strides, the fuzziness in their facial characteristics impedes accurate identification, culminating in misjudgment.

In terms of the overall experiment results, cloud computing demonstrates a significantly high level of accuracy in personnel recognition. Under the IoU (Intersection over Union) threshold of 0.5, an accuracy rate of approximately 83% was achieved. Furthermore, with robust GPU hardware support, cloud computing exhibits rapid inference times. For instance, in this experiment, the Tesla T4 GPU [40] available on Colab Pro was tested, yielding a prediction time of around 10 to 11 milliseconds (ms) per frame. This translates to a frame processing rate of approximately 95 frames per second (FPS), calculated as 100 milliseconds divided by 10.5 milliseconds. The FPS metric is a crucial measure of system processing speed. In our experiment, cloud computing showcases exceptional performance.

Figure 14.

Accuracy test result data chart.

Figure 15.

Misidentification problem.

5.3. Personnel recognition accuracy and inference time based on embedded AI

In this study, we aimed to evaluate the accuracy and response time of personnel recognition using embedded AI. To achieve this, we replaced the conventional approach of transmitting images from multiple Raspberry Pi devices to a server for processing with the use of Nano development boards. The Jetson Nano board was employed to perform real-time image recognition on the spot. To verify the personnel recognition accuracy in this research, we conducted experiments with real-time images similar to those used in the previous experiment (Section 5.2). As Table 9, the results showed an accuracy level of approximately 84%. Regarding Inference time, the Jetson Nano achieved a processing speed of about 9.5 frames per second (FPS) in its built-in environment. Additionally, we installed TensorRT, a software library developed by NVIDIA to accelerate deep learning inference. It is specifically designed for their GPU architecture. TensorRT focuses on optimizing models, providing low-level hardware acceleration, and achieving efficient deep learning inference. This library is widely utilized in deep learning inference applications on NVIDIA GPUs, including image recognition on embedded systems such as the Jetson Nano development board. TensorRT offers several key features. First, it enhances prediction speed by optimizing deep learning models for faster hardware execution. This is crucial for image recognition as faster prediction speed enables real-time object detection and classification. With TensorRT installed, the Jetson Nano development board performs image recognition with reduced latency, providing more immediate results. Second, TensorRT optimizes hardware resource utilization, making the most of the GPU and Tensor Cores in the Jetson Nano development board. This means that during the inference process, GPU and other hardware resources are utilized more efficiently, resulting in higher computational performance. This is especially important for image recognition on resource-constrained embedded systems, as it fully harnesses the potential of the hardware. Third, TensorRT allows for model compression, optimizing and reducing the size and memory footprint of deep learning models. This is particularly valuable for the Jetson Nano development board, which has limited storage and memory capacity. After conducting experiments, the original response time was improved from 9 FPS to approximately 16 to 17 FPS when TensorRT was added. This demonstrates that installing TensorRT on the Jetson Nano development board significantly enhances the speed and efficiency of image recognition, while optimizing the utilization of hardware resources. Consequently, it enables efficient image recognition applications in resource-constrained scenarios.

Table 9
Personnel identification accuracy based on embedded AI.

Experiment no. Region no. Test scenario User1 User2 User3 User4

1 A Corridor 66% 73% 64% 61%

2 B Corridor-corner 78% 90% 84% 78%

3 C Lobby 100% 100% 75% 77%

4 D Inside the elevator 92% 100% 100% 92%

Experiment no.	Region no.	Test scenario	User1	User2	User3	User4
1	A	Corridor	66%	73%	64%	61%
2	B	Corridor-corner	78%	90%	84%	78%
3	C	Lobby	100%	100%	75%	77%
4	D	Inside the elevator	92%	100%	100%	92%

Figure 16.

The main reason for the decrease in accuracy.

5.4. Comparison of cloud computing and embedded AI computing

In this section, we compare the performance of two architectures: cloud computing based and embedded AI based, based on the experiments conducted in Sections 5.2 and 5.3, respectively. Firstly, as Figure 16 in the experiments, the accuracy of both architectures is similar, with manual comparison results showing 83% and 84% for cloud computing and embedded AI, respectively. Normally, the Tiny model suitable for embedded systems would be expected to have lower performance compared to the original model. However, in this paper, we employed various optimization techniques, such as multi-scale training, increased training epochs, and image size adjustments, to enhance the average precision (mAP) of the model. As a result of applying these optimization methods, the performance of the Tiny model in the embedded system approached that of the original model in terms of accuracy.

Upon further analysis of the accuracy of cloud computing and embedded systems, we found that distance is the most critical factor influencing accuracy, as illustrated in Figure 13 and 16. Therefore, in the experimental process, we made slight adjustments to the scope of corridor recognition. If images beyond the region A need to be recognized, the accuracy will drop below 60%, which is far from the desired ideal state. Consequently, in the future, if recognition at longer distances is required, it is necessary to consider adding a camera in that area to improve the recognition rate.

In terms of inference time, as indicated by Table 10, we define three configurations: ConfigI, involving embedded AI combined with Raspberry Pi; ConfigII, involving embedded AI combined with Nano; and ConfigIII, involving cloud computation combined with Raspberry Pi. The cloud-based architecture poses a significant issue, as depicted in the rightmost column of Table 10. During the experimental process of transmitting images using multiple Raspberry Pi cameras, we observed that increasing the number of Raspberry Pi devices leads to a substantial influx of image data to the server. Apart from challenging network resources, this also proportionally extends the server’s inference time. In order to observe more conspicuously, ConfigIII experiment was conducted in a laboratory server without GPU. A single Raspberry Pi requires approximately 600 to 700 milliseconds to infer an image, equivalent to around 1.5 frames per second. Two Raspberry Pi devices require about 1200 to 1300 milliseconds to infer an image, equivalent to roughly 0.75 frames per second. Three or more Raspberry Pi devices will result in a linear reduction in inference speed. This implies that a cloud based architecture poses significant challenges to server performance. Even with well-equipped servers possessing GPUs and efficient network transmission environments, practical scenarios necessitate a substantial amount of camera equipment transmitting data in the vicinity of each floor. In the case of higher floors, this will inevitably lead to delays, potentially even reaching limits. Contrastingly, within the embedded AI architecture, each embedded camera device shares the server workload, as depicted in the ConfigI and ConfigII of Table 10. Consequently, this study posits that the embedded AI approach surpasses cloud computation in terms of superiority and stability. Within the embedded AI architecture, the server merely needs to receive reservation tag messages returned by the Nano board, without the need to transmit surplus image data. In addition, the experimental results demonstrate a significant disparity in processing speed between ConfigI, which operates at 0.8 FPS, and ConfigII, which achieves 17 FPS. This contrast highlights the substantial performance gap associated with the powerful GPU when dedicated image processing is involved. The inference speed of YOLOv7 Tiny with TensorRT technology has been proven to be sufficient to meet real-time image processing requirements.

Table 10
Inference speed comparison table.

Number of cameras Config I Config II Config III

1 webcam 0.8FPS 17FPS 1.5FPS

2 webcam 0.8FPS 17FPS 0.75FPS

3 webcam 0.8FPS 17FPS 0.375FPS

Number of cameras	Config I	Config II	Config III
1 webcam	0.8FPS	17FPS	1.5FPS
2 webcam	0.8FPS	17FPS	0.75FPS
3 webcam	0.8FPS	17FPS	0.375FPS

5.5. Elevator scheduling performance comparison

In order to compare the efficiency of elevator scheduling, this study chose the three most common elevator scheduling algorithms FCFS, SCAN, and LOOK to compare the LOOK+ proposed in this study for experimentation. In order to facilitate the observation of the differences between the algorithms, the experiment uses FCFS as the benchmark, and the percentage of waiting time that other algorithms can reduce compared with FCFS is the experimental result. In order to demonstrate the benefits of Grouping, LOOK+ was divided into two parts in the experiment. One is to remove Grouping, and the other is to include Grouping. These elevator scheduling algorithms have different strategies and priorities in handling passenger calls and have different effects on elevator operation. By comparing the average waiting time, average moving distance and other indicators of these algorithms in different scenarios, we can evaluate their performance differences.

Following the introduction of the algorithm types, the subsequent explanation will focus on the parameters utilized in the experiments. In order to ensure fairness in the simulations, identical variables are employed for all four algorithms. These parameters are as follows. Maximum floor number: 20. Number of elevators: 1. Time required to traverse one floor: 2 seconds. Initial starting position of the elevator: Floor 1. The task numbers (N), starting floors ( $C F$ ), destination floors ( $D F$ ), and waiting times ( $W T$ ) are generated by random numbers within a range smaller than the highest floor number. The starting and destination floors are different from each other. The waiting time, as shown in Table 11, refers to the time starting from when the passenger successfully reports to the elevator, with a maximum value of 300 seconds. The number of passengers is adjusted to mimic real-life scenarios for off-peak hours, regular hours, and peak hours. Peak Hours refer to periods within a day or a specific time frame when traffic volume is congested. This typically includes rush hours when people are commuting to and from work, students going to and from school, and busy commercial activities. Off-Peak Hours, on the other hand, are periods within a day or a specific time frame when traffic volume is relatively lower. Since there are no standard numbers for off-peak hours, regular hours, and peak hours, this experiment assumes passenger distribution based on common observations and typical scenarios. For peak hours, it is assumed that there are an average of 10 people waiting to take the elevator on each floor, resulting in a total of 200 people waiting. For off-peak hours, the assumption is an average of 2 people waiting on each floor, totaling 40 people. During regular hours, it is assumed that there are an average of 5 people waiting on each floor, totaling 100 people.

Table 11
Randomly generated simulation data illustration.

Request serial number Call floor Destination floor Waiting time

1∼N 1∼20 1∼20 0∼300

Request serial number	Call floor	Destination floor	Waiting time
1∼N	1∼20	1∼20	0∼300

This paper conducts experimental design for some consideration points. The first point is to consider the situation of different passenger flow. In the case of low load, because the passenger demand is small, the difference in the algorithm may not be obvious. Conversely, in extremely high load situations, where there is a large passenger demand, the effectiveness of the algorithm becomes more important. In this case, we can observe the impact of different algorithms and the algorithm proposed in this paper on the waiting time. The second point considers the number of different elevator configurations. If the elevator system has multiple elevators, we can observe whether the Grouping concept has a significant effect under this configuration. The third point is to consider the influence of different floors. A higher floor height may cause the elevator to move longer, which in turn will greatly affect the change in waiting time. In the analysis, it can be observed whether there is a situation where the higher the height is more sensitive to the effect of the algorithm. According to the experimental considerations of the above points, this experiment is designed to analyze the effect of the number of elevators and the number of floors on the waiting time of the algorithm under different traffic conditions.

Figure 17.

Simulation results for single elevator.

The first experiment is conducted in a single elevator environment. The results, shown in Figure 17, demonstrate that the SSTF algorithm achieves relatively lower waiting times compared to the other three algorithms. It optimizes waiting time by only 12%, 16.2%, and 18.4% in scenarios with 40, 100, and 200 people, respectively. However, the SSTF algorithm may also face issues in certain situations. For instance, if calls from a particular floor are consistently ignored, it can result in prolonged waiting times, causing inconvenience and unfairness. The LOOK algorithm’s primary advantage lies in reducing travel distances while also fairly distributing services. As the algorithm considers the elevator’s movement direction, it avoids excessive concentration on specific floors or directions, thereby reducing waiting times and travel distances. However, the LOOK algorithm also presents some problems, such as potential longer waiting times for certain passengers if the call distribution is uneven, affecting service quality. In comparison to the SSTF and LOOK algorithms, the proposed LOOK+ algorithm in this study demonstrates better performance. Specifically, when a new call is received, the algorithm automatically adjusts the elevator’s scheduling based on the waiting time. If a call has a longer waiting time, the algorithm prioritizes it, even if it may not be the optimal solution in terms of travel distance. This ensures that calls with longer waiting times receive timely attention. However, in the single elevator simulation, the performance difference between the proposed algorithm and the LOOK algorithm is not significant. This is because the LOOK algorithm already has a certain level of optimization, limiting the improvement potential of the proposed algorithm in such a simple single elevator environment.

In the second experiment, this experiment further expanded the number of elevators and conducted experiments using five elevators in a building with 20 floors. The number of elevator calls was set to 40, 100, and 200 people. The experimental results are shown in Figure 18. The proportions of the first three algorithms are similar to those in the first experiment. This is because as the number of elevators increases, the tasks are evenly distributed among each elevator, leading to a proportional decrease in the total waiting time. Thus, the percentage reduction in waiting time is similar among these algorithms.However, the proposed LOOK+ algorithm exhibits different performance. According to the analysis in this study, Grouping, which groups passengers together, directly influences the task allocation in elevator scheduling. Processing similar tasks together reduces the number of elevator stops at a particular floor, leading to improved efficiency. Particularly, when the number of elevator calls is 200 people, the average waiting time achieved by the LOOK+ algorithm is optimized to 32.2%, significantly outperforming the LOOK algorithm’s 25.9% and the SSTF algorithm’s 18.4%, demonstrating superior performance.

Figure 18.

Simulation results of multi-elevator.

Finally, this experiment further investigates the impact of building height on the effectiveness of the algorithm. We set the building heights to 10 floors, 30 floors, and 50 floors, respectively, and conducted experiments with a call volume of 100 people during normal periods. The experimental results are shown in Figure 19. Based on the observed results, the average waiting time for all four algorithms shows a clear increasing trend with varying building heights. This is because as the building height increases, passenger demand is more widely distributed, and elevators require more time for vertical movement, resulting in increased waiting times. However, the proposed algorithm in this study maintains a better scheduling performance at different building heights, reducing waiting times by 16.1%, 32.9%, and 37.2% for 10 floors, 30 floors, and 50 floors, respectively. This indicates that regardless of building height, the algorithm proposed in this study effectively reduces passenger waiting time, with a more significant improvement at higher heights. These results demonstrate the feasibility and superiority of the proposed algorithm in practical applications, making it valuable for elevator systems in buildings of various heights.

Figure 19.

Simulation results of building height.

To summarize, this experiment compared the effects of different algorithms and the concept of Grouping on waiting times under various scenarios, including different crowd flows, elevator configurations, and floor heights. The results demonstrate that the proposed algorithm in this study performs remarkably well across all scenarios. In the context of multi-elevator setups, it can significantly reduce passenger waiting times by at least approximately 30%, with an even more pronounced effect as the number of floors increases, thus enhancing the overall efficiency of the elevator system. These findings provide valuable insights and guidance for the design and optimization of elevator systems.

6. Conclusion

In this paper, an intelligent reservation and scheduling system for elevators – SmartRide is proposed. The system incorporates a comprehensive reservation mechanism and enhances the most prevalent elevator algorithms in use today. Users can easily make elevator reservations through the provided interface. Additionally, the system employs embedded systems combined with image recognition technology for personnel identification, enriching the algorithm with more comprehensive user data. Furthermore, our algorithm enhancement is based on the widely used LOOK algorithm, combining factors such as passenger booking information, current elevator status, and waiting times to optimize scheduling. This enhancement effectively reduces average passenger wait times, thereby enhancing user experience. In the experimental results, we compared different personnel identification methods, achieving accuracy rates exceeding 80% in both cloud computing and embedded AI architectures. Notably, the embedded AI architecture exhibited more consistent response times. Moreover, in terms of algorithms, we simulated and compared the SSTF algorithm, LOOK algorithm, and our proposed LOOK+ algorithm against the traditional FCFS algorithm. The results demonstrated that in various high-rise scenarios, our approach consistently reduced average wait times by at least 30%, with even more significant improvements as the number of floors increased. These findings affirm the superiority of our research method in enhancing overall system efficiency and call satisfaction.

Through the integration of edge computing technology, our study presents an innovative solution that leverages intelligent recognition systems and algorithms to optimize elevator scheduling, ultimately improving passenger wait times. This research contributes to enhancing elevator system performance and passenger satisfaction in modern urban settings, aligning with the development of future smart cities. Going forward, we will continue refining the user interface to enhance application usability. Simultaneously, we will investigate more intelligent algorithms to handle scenarios such as non-reserved passengers’ calls and complex situations where passengers without reservations also require elevator access. Beyond interface and algorithms, we will persist in enhancing system stability and security to ensure the protection of user information. Additionally, we aim to apply the system in specific contexts with unique elevator demands, such as hospital transport personnel, by enabling advanced elevator reservations to further enhance operational efficiency.

Footnotes

Acknowledgements

The authors express their gratitude to Tsu-Chuan Shen, Chang-Yu Ling, Kai-Di Zhang, and Jin-Wei Hou for their valuable assistance in the data collection process for the experiment. Additionally, appreciation is extended to ChatGPT for its contributions in checking the English grammar. Recognition is also given to the online icon database of Freepik Company (https://www.flaticon.com/) for generously providing researchers with icons free of charge.

Conflict of interest

The authors have no conflict of interest to report.

References

Al-Kodmany

, The vertical farm: A review of developments and implications for the vertical city, Buildings 8(2) (2018), https://www-mdpi-com-s.web.bisu.edu.cn/2075-5309/8/2/24. doi:https://doi.org/10.3390/buildings8020024.

Artificial lift systems market size, share & trends analysis report by product (rod lifts, electrical submersible pumps), by region, and segment forecasts, 2018–2025.

Houston

M.B.

Bettencourt

L.A.

Wenger

, The relationship between waiting in a service queue and evaluations of service quality: A field theory perspective, Psychology & Marketing 15(8) (1998), 735–753. doi:https://doi.org/10.1002/(SICI)1520-6793(199812)15:8<735::AID-MAR2>3.0.CO;2-9.

Norman

D.A.

, The psychology of waiting lines, Excerpt of 3, 2008.

Caruelle

Lervik-Olsen

Gustafsson

, The clock is ticking – or is it? Customer satisfaction response to waiting shorter vs. longer than expected during a service encounter, Journal of Retailing (2023).

Kropp

Faeghi

Lennerts

, Evaluation of patient transport service in hospitals using process mining methods: Patients’ perspective, The International Journal of Health Planning and Management 38(2) (2023), 430–456. doi:https://doi.org/10.1002/hpm.3593.

Shen

T.-C.

Chu

E.T.-H.

, Edge-computing-based people-counting system for elevators using MobileNet–single-stage object detection, Future Internet 15(10) (2023), 337. doi:https://doi.org/10.3390/fi15100337.

New study shows modern working patterns are increasing lift waiting times and carbon emissions.

Mossberg

Nilsson

Andrée

, Unannounced evacuation experiment in a high-rise hotel building with evacuation elevators: A study of evacuation behaviour using eye-tracking, Fire technology 57 (2021), 1259–1281. doi:https://doi.org/10.1007/s10694-020-01046-1.

10.

Chen

, Improvement of hospital operation elevator dispatching system, PhD thesis, ETSI_Informatica, 2021.

11.

Akar

Tunç

Ç.

Gaertner

Erden

, Disk scheduling with shortest cumulative access time first algorithms, Turkish Journal of Electrical Engineering and Computer Sciences 25(4) (2017), 3367–3380. doi:https://doi.org/10.3906/elk-1610-271.

12.

Gudwin

Gomide

Netto

M.A.

, A fuzzy elevator group controller with linear context adaptation, in: 1998 IEEE International Conference on Fuzzy Systems Proceedings. IEEE World Congress on Computational Intelligence (Cat. No. 98CH36228), Vol. 1, IEEE, 1998, pp. 481–486.

13.

Kim

C.B.

Seong

K.A.

Lee-Kwang

Kim

J.O.

, Design and implementation of FEGCS: Fuzzy Elevator Group Control System, in: Proceedings of North American Fuzzy Information Processing, IEEE, 1996, pp. 109–113. doi:https://doi.org/10.1109/NAFIPS.1996.534713.

14.

Fernández

Cortés

Muñuzuri

Guadix

, Dynamic fuzzy logic elevator group control system with relative waiting time consideration, IEEE transactions on industrial electronics 61(9) (2013), 4912–4919. doi:https://doi.org/10.1109/TIE.2013.2289867.

15.

Fujino

Tobita

Segawa

Yoneda

Togawa

, An elevator group control system with floor-attribute control method and system optimization using genetic algorithms, IEEE Transactions on Industrial Electronics 44(4) (1997), 546–552. doi:https://doi.org/10.1109/41.605632.

16.

Cortés

Larrañeta

Onieva

, Genetic algorithm for controllers in elevator groups: Analysis and simulation during lunchpeak traffic, Applied Soft Computing 4(2) (2004), 159–174. doi:https://doi.org/10.1016/j.asoc.2003.11.002.

17.

Liu

Huo

, A multi-objective genetic algorithm designed for energy saving of the elevator system with complete information, in: 2010 IEEE International Energy Conference, IEEE, 2010, pp. 126–130. doi:https://doi.org/10.1109/ENERGYCON.2010.5771661.

18.

Tartan

E.O.

Erdem

Berkol

, Optimization of waiting and journey time in group elevator system using genetic algorithm, in: 2014 IEEE International Symposium on Innovations in Intelligent Systems and Applications (INISTA) Proceedings, IEEE, 2014, pp. 361–367. doi:https://doi.org/10.1109/INISTA.2014.6873645.

19.

Erdem Imrak

Özkirim

, Determination of the next stopping floor in elevator traffic control by means of neural networks, IU-Journal of Electrical & Electronics Engineering 6(1) (2006), 27–33.

20.

Crites

R.H.

Barto

A.G.

, Elevator group control using multiple reinforcement learning agents, Machine learning 33 (1998), 235–262. doi:https://doi.org/10.1023/A:1007518724497.

21.

Ikuta

Takahashi

Inaba

, Strategy selection by reinforcement learning for multi-car elevator systems, in: 2013 IEEE International Conference on Systems, Man, and Cybernetics, IEEE, 2013, pp. 2479–2484. doi:https://doi.org/10.1109/SMC.2013.423.

22.

Strang

Bauer

, Context-aware elevator scheduling, in: 21st International Conference on Advanced Information Networking and Applications Workshops (AINAW’07), Vol. 2, IEEE, 2007, pp. 276–281. doi:https://doi.org/10.1109/AINAW.2007.131.

23.

Amano

Yamazaki

Ikejima

, The latest elevator group supervisory control system, in: Elevator Technology 6 Proceedings of Elevcon, Vol. 6, 1995, p. 88.

24.

Hangli

Hamada

Sumitomo

Koshizuka

, Intellevator: An intelligent elevator system proactive in traffic control for time-efficiency improvement, IEEE Access 8 (2020), 35535–35545. doi:https://doi.org/10.1109/ACCESS.2020.2975020.

25.

Chen

Zheng

Liu

Ren

, A real-time matrix iterative optimization algorithm of booking elevator group and numerical simulation formed by multi-sensor combination, Electronics 10(24) (2021), 3179. doi:https://doi.org/10.3390/electronics10243179.

26.

Khonjun

Pitakaso

Sethanan

Nanthasamroeng

Pranet

Kaewta

Sangkaphet

, Differential evolution algorithm for optimizing the energy usage of Vertical Transportation in an Elevator (VTE), taking into consideration rush hour management and COVID-19 prevention, Sustainability 14(5) (2022), 2581. doi:https://doi.org/10.3390/su14052581.

27.

Van

L.-D.

Lin

Y.-B.

T.-H.

Lin

Y.-C.

, An intelligent elevator development and management system, IEEE Systems Journal 14(2) (2019), 3015–3026. doi:https://doi.org/10.1109/JSYST.2019.2919967.

28.

Cortés

Guadix

Muñuzuri

, A state of the art on the most relevant patents in vertical transportation in buildings, Recent Patents on Computer Science 2(2) (2009), 96–110.

29.

Turunen

Kuoppala

Kangas

Hella

Miettinen

Heimonen

Keskinen

Hakulinen

Raisamo

, Mobile interaction with elevators: Improving people flow in complex buildings, in: Proceedings of International Conference on Making Sense of Converging Media, 2013, pp. 43–50.

30.

Kwon

Lee

Bahn

, Sensor-aware elevator scheduling for smart building environments, Building and Environment 72 (2014), 332–342. doi:https://doi.org/10.1016/j.buildenv.2013.11.013.

31.

Homchalee

Pitakaso

Kumphon

, Minimizing energy usage and makespan of elevator operation in rush hour using multi-objective variable neighborhood strategy adaptive search with a mobile application, Mathematics 11(8) (2023), 1948.

32.

Andriana

, Elevator door automatic stopper for Covid-19 physical distancing, in: Proceedings of the 2021 International Conference on Computer, Control, Informatics and Its Applications, 2021, pp. 1–5.

33.

Lai

S.-C.

H.-H.

Hsu

W.-L.

Wang

R.-J.

Shiau

Y.-C.

M.-C.

Hsieh

H.-N.

, Contact-free operation of epidemic prevention elevator for buildings, Buildings 12(4) (2022), 411. doi:https://doi.org/10.3390/buildings12040411.

34.

Rubies

Bitriá

Clotet

Palacín

, Non-contact and non-intrusive add-on IoT device for wireless remote elevator control, Applied Sciences 13(6) (2023), 3971. doi:https://doi.org/10.3390/app13063971.

35.

Aljadani

S.M.

Almutairi

S.M.

Ghaleb

S.S.

Al Khuzayem

, Elevint: A cloud-based Internet of elevators, International Journal of Advanced Computer Science and Applications 13(1) (2022).

36.

Chandirasekeran

Shridevi

, Smart elevator obstruction detection system using image classification, International Journal of Advanced Computer Science and Applications 13(4) (2022).

37.

Zhang

Sun

Zhao

, Elevator ride comfort monitoring and evaluation using smartphones, Mechanical Systems and Signal Processing 105 (2018), 377–390. doi:https://doi.org/10.1016/j.ymssp.2017.12.005.

38.

Matsumoto

Yamazaki

Igarashi

Yoshida

Shiraishi

, Smart elevator Hall: Prototype of in-building guiding using interactive floor display, in: Companion Proceedings of the 2020 Conference on Interactive Surfaces and Spaces, 2020, pp. 15–18. doi:https://doi.org/10.1145/3380867.3426207.

39.

Wang

C.-Y.

Bochkovskiy

Liao

H.-Y.M.

, YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7464–7475.

40.

Google, Google colaboratory, accessed: August 13, 2023.

SmartRide: Intelligent reservation and scheduling for elevators

Abstract

Keywords

1. Introduction

1.1. Current situation and problem

1.2. Existing solutions of intelligent elevators

1.3. Our contributions

2. Related work

2.1. Elevator scheduling

3.1. Minimizing waiting time

3.2. Providing reservation and auto recognition

4. Methodology

4.1. System architecture

4.2.2. Check-in process

4.2.3. Boarding process

4.3. Personnel recognition method based on cloud computing

5.1. Experiment design

5.2. Personnel recognition accuracy and inference time based on cloud computing

Table 9 Personnel identification accuracy based on embedded AI. Experiment no. Region no. Test scenario User1 User2 User3 User4 1 A Corridor 66% 73% 64% 61% 2 B Corridor-corner 78% 90% 84% 78% 3 C Lobby 100% 100% 75% 77% 4 D Inside the elevator 92% 100% 100% 92%

Table 10 Inference speed comparison table. Number of cameras Config I Config II Config III 1 webcam 0.8FPS 17FPS 1.5FPS 2 webcam 0.8FPS 17FPS 0.75FPS 3 webcam 0.8FPS 17FPS 0.375FPS

Table 11 Randomly generated simulation data illustration. Request serial number Call floor Destination floor Waiting time 1∼N 1∼20 1∼20 0∼300

Footnotes

Acknowledgements

Conflict of interest

References

Table 9
Personnel identification accuracy based on embedded AI.

Experiment no. Region no. Test scenario User1 User2 User3 User4

1 A Corridor 66% 73% 64% 61%

2 B Corridor-corner 78% 90% 84% 78%

3 C Lobby 100% 100% 75% 77%

4 D Inside the elevator 92% 100% 100% 92%

Table 10
Inference speed comparison table.

Number of cameras Config I Config II Config III

1 webcam 0.8FPS 17FPS 1.5FPS

2 webcam 0.8FPS 17FPS 0.75FPS

3 webcam 0.8FPS 17FPS 0.375FPS

Table 11
Randomly generated simulation data illustration.

Request serial number Call floor Destination floor Waiting time

1∼N 1∼20 1∼20 0∼300