Adaptive path planning for unknown environment monitoring

1. Introduction

Nowadays, obtaining information is becoming easier, with the complexity of accessing data being comparable to that of accessing a memory block. However, getting data from a sensor network is critical. Currently, data is collected through sensor networks. In the physical realm, data is collected for inspecting oil pipeline’s leakage, reviewing infrastructure for defects, researching water bodies to monitor the environment, and scanning risk events for relief operations. The limitation of fuel usage, as well as the length and distance of trip for data gathering, are notable issues in data gathering. As a result, path planning is an NP-hard task. The current state of the art systems take two approaches to solving this challenge. A myopic technique [10,56] is one approach, in which a set of sensing locations is determined and the system travels to the most promising one. While such solutions are computationally efficient, they fail to account for the travel distance constraint properly.

Various types of robots become popular for general, medical, industry and agriculture purposes. Unknown environment path planning for an autonomous exploration and simultaneous localization for agriculture robot got increased consideration [34]. Highest accurate environmental map is basis for bionic robot [28], autonomous vehicle [15], unmanned aerial vehicle [14] and Agricultural robot [24]. These robots achieve navigation, picking and placing things, surveying. The real challenge is the changing environment, variation in shape, illumination, occlusion, view point, etc. it is difficult to create maps for every real time scene. The absence of intelligent exploration of environment is cutoff the habit of robot expertise [48]. Consequently numerous researches originated to train robot, independently explore path planning for unknown environment, which permit the robot to develop dynamic map [26]. Optimal path plan generated by which is observed locally is combined with a partially created map [39,40,45,46,54]. The robot simultaneously predicts the map and updates the map to navigate along this trajectory [29]. Many researchers focused about the exploration of sensor’s position optimization. Full coverage observation of environment perception is achieved by frontier-based approach [32]. Frontier-based approach cannot guarantee the optimal trajectory plan because of its single stage greedy method [58]. By increasing the information gained through investigation, another method named multi step optimization used acquire optimal path [22]. Information gathering again based on monitoring the environment persistently. In many applications, maintaining the changing parameter is a crucial for managing uncertainty in environment path planning [2–4,12,17,19,20,23,35,37,41,43,44].

The evolution of robotics-based monitoring has ushered in a slew of new research opportunities. The fact that real devices have restricted amount of perceiving sources, like power, duration, or distance traveled, limits the number of observations which can be obtained is a major obstacle. Therefore, path planning need to planned extremely to maximize the time, energy, fuel. Path Planning plays vital primitive against autonomous agriculture robots that can find the shortest or otherwise optimal path between two points on the agriculture field or farm. Path planning is a route finding problem of effective outline to move the robot from the source to target. When the environment is newer than the path planning requirements, the map of the field requires the robot to be intelligent about its position. There are many robots available for real time scenario such as painting robots, cleaner robots, mining robots, lawn mower robots [43], Sowing robot 1 [37], Lumai-5 [2], Di-Wheel [44], and AgBot [17]. Path planning in agriculture are extend in various application. Two types of path planning proposed for agriculture robot: the first is point to point based path planning, the second is coverage based path planning [35]. In non-adaptive context, when the robots executed plan a path without accounting for noisy observations, the informative path planning (IPP) complexity is NP hard. The adaptive setup increases the difficulty by needing not only a strategy, but also a policy that adapts to the observations the agent receives. We further expand the problem by explicitly introducing different sensing modalities and taking into account the impact of both sensing and moving on the energy budget [4].

Autonomous robots are collectively working to gain worthy data about our regions. In the older days, rapid technological advances have unlocked unique stages of autonomy and decision-making of their perception just in case of flexible and cost effective tool [20]. Strategies are sapping conventional data collecting procedures using fixed sensors, conventional sampling, or manned platforms that are steered in aquatic applications, this can be inconvenient, costly, and even deadly [23].

Rational devices focus a challenge that is limited size of sensing resources, for example distance, session, energy, which restricts the quantity of factors which can be gathered. Hence, path must be well planned, enlarge the information gather about the unknown environment, whereas fulfilling the set of budget constrain. This IPP issue, which is the focus of current research, despite ongoing improvements in this field, most of the robot acquired data in passive manner only. IPP using coverage-based path planning are usually restricted to a few platforms or application domains. Intelligent vehicles are an intelligent agent which consist of environment perception & modeling, localization with mapping, path planning with decision-making and motion control.

Mostly others have spent their effort by comparing algorithms in order to determine which can provide the most precise map. Every method has such a set of conditions that can be tailored to the turtle’s work atmosphere. Factors like robot rapidity, plotting apprise interval can affect not only how long it takes but also how accurate the map is. As a result, it is vital to understand how different algorithm settings modify map characteristic, which is the focus of this research, particularly when robots are operating in unfamiliar environments especially when robots are working in unknown environment.

Formerly, fuzzy logic and neural networks were used to produce a variety of control programmers, system identification techniques, and active disturbance rejection platforms. While neural networks have complex neuron controllers, designing a fuzzy logic controller requires expert knowledge. Although the Negative-Imaginary (NI) theory is applicable to robust vibration control issues for adaptable structures with high-resonant dynamics, most approaches based on this theory have not addressed the challenge of high-precision trajectory monitoring for quad rotor structures in environments that vary.

Machine learning strategies are widely used in autonomous robots in recent years. These robots are equipped with adaptive learning capabilities based on acquired knowledge. An essential subfield of machine learning is reinforcement learning (RL), which continually updates an agent’s activation policies in response to feedback dependent on the situation by trial-and-error. The availability of training samples has increased due to the advancement of simulation technologies and the accumulation of measurement data. Due to its strong ability to learn in challenging contexts, reinforcement learning is shown feasible outcomes in the navigation of mobile robot. The designing of track planning approach polished, then these algorithms parameter improved by valuable facts. Early research also employed ML to increase the environment awareness, which is huge, dynamic, or partially unstructured.

Also, manually tuning complex control loop requires more time to controller experts. Additionally, single-objective criteria like small chasing fault, short mounting time and minimum overshoot are satisfied via manual parameter tuning and optimization. Therefore, intelligent embedding method is presented to deal these issues. The global optimization strategies for parameter tuning contain genetic algorithm, ant colony optimization, particle swarm optimization, reinforcement learning etc. Due to comprehensive onboard implementation, Machine Learning strategies allow automatic adaptation to environment. But define the performance of potential tuning parameter solutions and gather a big size of information through relevant real-world domain, its evaluation framework necessitates a significant amount of time. There are other sources of uncertainty that could contribute to the multiple disruptions in performance evaluation (sensor noise or manufacturing tolerances). Significant research theme at robotics is the higher-performance route track of unmanned aerial vehicle by effect of aging as well as environment. Its advantages contains a broad spectrum of application in civilian and military domain, like aircraft review, following, implementation of law, wildlife observing. If robust systems of control are not designed to withstand the impacts of external commands and disturbances, this sort of uncertainty results in poor performance or instability in the system. Despite the fact that numerous path-following control strategies have been investigated to mitigate the impact of the additional wind component, there is a high reliance on wind knowledge and a standard PID monitoring controller architecture with no active anti-disturbance features capacity compared to other tuning models.

As already indicated, the majority of existing path planning approaches in use mainly relies on a static global map. When the map contains vibrant restrictions, its state requires to be recognized. These algorithms may act badly with unknown static with dynamic obstacles and frequently get orientation track instead of velocity facility, so extra sensors or tracking techniques must be used, which complicates navigation along computation for the mobile robot.

This is the reason behind the motivation of this research. To overcome the shortage of autonomous vehicle especially for unknown environment depends on accurate map to achieve the effective path planning and improve optimization for agriculture phenotype robot and enhance the persistent monitoring based on region of interest. The key contribution from our side work is a HSAStar mapping approach for point cloud data is utilized to calculate the various spatial arrangements for unknown environment. Then save them in the map to promote persistent monitoring. Hybrid SLAM & A Star algorithm for path planning is convenient to activate with the parameters for tuning. It proves a greater convergence, lesser computation time, and better proficiency to determine global optimum. Obviously, its recall ability is more effectual than learning algorithms, because every star can able to concurrently signify own preceding better place and neighborhood better.

2. Problem statement

For a robot travelling across a landscape, our design stresses data-gathering strategies. The goal is to gather as much environmental data as possible while remaining within power, spell restrictions and distance costs are examples of resource constraints. This problem is properly called as path planning for efficient data gathering and is defined as follows. This section explains how to formulate the Persistent Monitoring Problem it is frequently used to represent persistent surveillance problems by dividing the space into separate cells with a measurement [16,57] are concerned with the effects of monitoring. We make use of the graph to develop the discrete model, use theory as a simplification technique of the actual path formation, which is made up of perspectives as well as the edges. UGV has a monitoring influence on the viewpoints. Sensor detection ability was used to calculate the value based on the PMP formulation form is supplied on the above. The purpose of this paper is to show how a UGV may cover a large unknown a large region in a short period of time while detecting a large number of events. We’ve seen events that we’re concerned about out of nowhere in an unknown setting. The likelihood of their occurrence as well as the regions of their occurrence are unclear. Although the exact time of an event’s commencement is unknown, its length can be calculated statistically [55]. If the UGV accurately detects a continuing event, it signifies the UGV has fruitfully record the event’s incidence in an unknown environment [13]. To put it another way, if the incident lasts longer than the spell between two exact UGV monitoring excursions on same point, it can be located at any time. Finding UGV trajectories with a specific coverage limit while they are moving in order to maximize a scrutinizing norms related to the observing effect. Priorities, based on the likelihood of an event occurring and its location are a mystery.

The PMP task environment model is abstracted into a node represents connected Graph = ( V , E ) , which is comprised of a network of discrete points of view and straight edges, wherever V = v 1 , … , v N is a view point that UGV must monitor, N represents of points of view, and E = e 1 , … , e M is the edge that UGV need travel, and M is the total of edges. | E a , b | denotes the span of | E a , b | , whereas E a , b  E denotes edge linking views v a and v b . When examining spatial attribute of occurrence, length of a cutting edge in graph is not huger than the frequency scope of the minimal occurrence. The occurrences with the minimal effect range must be considered noticed in the task of monitoring, for plant leaf spot disease identification, the frequency range of minimum event is 1 to 3 meter in diameter. As a result, when UGV keeps a close eye on the situation nearby perspectives on it, it is able to discover activities stirring on the edge. By calculating the monitoring the impact of distinct perspectives, unknown environment monitoring effect may be determined. The phrases “known” and “unknown” apply to an agent’s state of knowledge when executing an action, not to elements of the environment. In a known environment, the agent knows the consequences of all actions. To perform an action in a new environment, the agent must first understand how it works. A known environment may be somewhat visible, but an unfamiliar setting may be completely visible. In our research the unknown actions are, robot cannot segment the agriculture field which unfamiliar to the robot as semantically after its reconstruction. Second it don’t known about what action need to be done for navigation from one place to another.

In this, the uncertainty is a problem reduced by assuming the robot, the relative pose of teammate ca be derived from relative distance ρ (presented via laser), the angle β j i wherein the robot i, the angle β i j at which robot j, observes robot i (angular measurements simply acted by cameras). Here, the uncertainty problem is observed by assuming an angle β i j with final odometric pose of robot j at the robot R jo is formerly interacting with robot j. The ρ distance is scaled by i robot, therefore ρ = ρ j i . If ρ j i communicate by robot j, even though not required, it can be averaged with ρ j i allowing, under the hypothesis of independent Gaussian noise, to lessen the distance measurement variance, the uncertainity is expressed in equation (1), (1) σ j 2 = σ ρ j i 2 σ ρ j j 2 2 / ( σ ρ j i 2 + σ ρ j j 2 2 ) From an initialization agriculture land graph representation of a model G = ( Vertex , Edge ) we will now go over how to generate a stochastic graph representation of a model h r ( R ) is given in equation (2) (2) h r ( R ) = ∑ i = 1 a h i ( Ri , 1 RIi , 2 ) In equation (1), ∑ i = 1 a h i ( Ri , 1 R Ii , 2 ) is weight function ( W i ’), h i is represented as the discrete contour, a is the best approximation, Ri is the R i = ( r 0 ; r 1 ; … ; r n ) is represented as the resampled contour, Ii is represented as the graph co-ordinates, PMP is represented as d e , and the permitted modelling deviation in (3) (3) PMP ( W t ) = H ( W ∗ ( R ) ) here e ( 0 , 1 ) has been the initial condition coefficient of precision discrete, which is defined in equation (3) by the discrete precision of the unknown environment path. The ith path is designated by W i belongs to W, W t = { w t 1 , … , w t a } where a is the amount of path in agriculture field. R = { R 1 , … , R b } are path endpoints, b – intersections, R ( i , 1 ) and R ( i , 2 ) are the points where the ith path W i intersects with W i , and ( R i , 1 , R i , 2 ) belongs toI.

Step 1: Path quantization

OPEN IN VIEWER

Fig. 1.

Discrete process of path.

Step 2: Long Path Segment

Any road Wt i N in W N is labeled as elongated line if | Wt i N | > d e , where | Wt i N | specifies the length of Wt i N . To satisfy | Wt i N | l s ⩽ d e , find an integer s, and the elongated, the line would be broken into s parts. Add all recent nodes and segments as of previous stages to I, and delete W i N from W N , decisions on whether or not to divide the path will be made again and again until there are no lengthy queues in W t N .

Step 3: Creation of Grid Representation

Entire intersections created through preceding phases are referred to as views and are denoted by the letters V = { v 1 , … , v N } . Two perspectives joined directly to formulate an edge based on the topology path. E a , b ∈ E is the edge that connects the views v a and v b .

2.1. Problem definition and assumptions

Likewise the above mentioned task of warehouse, the uncertainty issue is solved by modifying the motion parameter and lowering velocity prior to the start of the study in order to avoid the motion route from diverging from the planned path due to the too high velocity at the entrance and exit points. The robot’s initial position and target are pre-set, the robot recognizes the global map outline (walls of warehouse), i.e, the problems in map are entirely unidentified. The aim of the robot is to reach the destination while keeping a safe distance from the obstacles. The robot can reach the destination at any velocity and does not need to stop.

In addition to the global map’s layout, robot’s many sensors provide all the information it can (gyroscope, lidar, encoder). These sensors can be used to gather the data needed by the proposed method at each time step. Make given assumption regarding navigation issue.

Assumption 1.

The robots place, alignment and velocity at the global map identified in every time step.

Assumption 2.

The target place in the global map identified in every time step.

Assumption 3.

The closest obstacle distance in each circle direction is identified at every time step. These data structure is exhibited as [ ⊗ d ′ ] , here ⊗ = [ ϕ 1 ϕ 2 … ϕ n ] T ′ are angle in range [ 0 , 2 π ] with resolution; d ′ = [ D 1 ′ D 2 ′ … D n ′ ] T ′ are distance of nearest problem in ith angle, D j is not better than discovery range; n indicates lidar angle resolution, which implies that angle and distance data pairs will be collected once the lidar scans a circle.

The robot is differentially propelled and considers itself to be moving forward along a rotating axis. Equation (4) depicts the robot’s kinematics theory. (4) a = ϑ cos θ b = ϑ sin θ θ = ϖ ϑ = x ϖ = β where [ a , b ] represents the robot position; θ represents robot orientation; ϑ and ϖ represents robot’s linear velocity with angular velocity; x and β represents robot’s linear acceleration with angular acceleration. Then the parameters, such as ϑ, ϖ, x and β represents optimization parameter for optimally reduce the uncertainty problem, such as robot’s linear and angular velocity and robot’s linear and angular acceleration. These parameters are optimized using A star algorithm.

3. Method proposed for path planning under an unknown environmental using hybrid Simultaneous Localization and Mapping (SLAM) with A Star Algorithm

3.1. Simultaneous Localization and Mapping (SLAM):

SLAM allows the rapid creation of navigable outlines that converts as navigable spaces and maps. This is impractical to use manual methods for mapping indoor environments, because of complexity and size of indoor space. Traditional SLAM has been used primarily in the field of robot navigation, requiring high level of output precision in order to be useful for robot movement maps. This has prompted research on more effective and quick sensors, measurement techniques, and measurement data post-processing. However, every improvement in accuracy has been accompanied by an increase in processing or equipment costs.

The Robot Operating System (ROS) platform, which is existing in Ubuntu, may be used to do SLAM, It is among the approaches for mapping robots on the move. GMapping is used as the algorithm for this study. The responsibilities of cartography, positioning, target tracking tasks, overlapping hitches. The difficulty of constructing a map while simultaneously localizing a robot within that map is known as SLAM. Both tasks can’t be decoupled and solved separately. Localization requires a good map, whereas building a map necessitates an accurate pose approximation. To improve the pose estimation, the goal of effective localize is to guide the robot to specific locations on the grid. Exploration approaches, on the other hand, presume precise posture knowledge and concentration on efficiently directing the vehicle across the surroundings in order to develop a map.

OPEN IN VIEWER

Fig. 2.

Key component process of the robot.

Solutions to the SLAM problem are sometimes known as integrated techniques. The key component process of the robot using SLAM algorithm is shown in Fig. 2. The Rao-Blackwell zed Particle Filter (RBPF) is utilized by G Mapping to collect data laser sensors and other sources. Make a 2D grid map with a robot stance [50]. In essence, the fundamental issues in map learning are listed as what place to steer a robot on a mission of self-exploration, how to handle noise in posture estimate and observations, and how to manage with ambiguity and how to interpret it data from sensors, how to effectively manage a team of mobile robots and variations in the echo system over time [59]. The mapping flow procedure is depicted in Fig. 3.

OPEN IN VIEWER

Fig. 3.

Environment mapping flow process.

There are a number of advantages to using GMapping as the primary HSAStar technique in this work. For starters, it’s an open source, algorithm which is ready to use, that’s part of the Robot Operating System (ROS). Second, it is deep rooted and has a proven track record, as evidenced by recent publications such as [31]. Finally, because the approach is a solo threaded technique in which each element is calculated separately, it lends itself nicely to parallelization. This is a good situation named at parallelization because there is no want to transfer data to particles. Whenever the robot acquires a fresh laser shot, the program executes steps outlined in the flowchart in Fig. 3. The vehicle alters the posture of the whole examined particle throughout execution based on odometry estimate using its motion model. The first scan is immediately recorded in the map when it is received. After that, only if the robot’s linear or angular distance traversed exceeds specific thresholds is registration performed. After that, data acquisition match is performed, which corrects pose estimation in each particle’s map. As a result, the particle tree’s weights have been adjusted, as has the effective sample size P ef is used to determine whether the present particle set performs posteriorly resembles goal particle set. (5) P ef = ( 1 − ω ) h ( n ) + ω l ( n ) Finally, equation (5) represents ω l is the normalized weight of particle i. whenever P ef is less than N / 2 , where number of particles represented by N, l ( n ) of l ∗ ( n ) then fulfills l ( n ) ⩾ l ∗ ( n ) . Likewise, h ( n ) is evaluated with evaluation function h ∗ ( n ) . The resampling stage is performed on the number of elements, because of the number of resampling processes. While a run, the amount of time spent on it is reduced, and it is only used when absolutely necessary. The reader should see the Gmapping method for more information [49].

For posture estimation and map generation, this study uses the data-matching approach then the grid representation using Gmapping algorithm. As illustrated in (4), HSAStar algorithms are frequently built on a common filter using probabilistic Bayes, with specific data being used to assess the density of an unknown probability density. The Bayes Filter, also known as Recursive Bayesian Estimation, is the foundation of several probabilistic techniques in robotics. A state, such as a robot stance, is represented by a Probability Density Function, or pdf, in Probabilistic Robotics. Another term is belief, which refers to the robot’s internal knowledge, rather than ours, from its perspective. The Bayes Filter is a general method for iteratively estimating a state’s pdf using a Mathematical Process Model and Measurements. (6) P ( x ) = ∑ i = 1 N w i δ s ( x ) The procedure, which is recursive, calculates the belief distribution from the previous belief distribution. The algorithm consists of two steps, one is to make a prediction based on the prior state and control data. The distribution based on forecast and the measurement data during the measurement update which second stop.

3.2. A Star algorithm for point cloud data

A Star is equivalent to Greedy Better in that it uses a heuristic to direct itself. In the simplest instance, it’s as speedy as Greedy Best-First-Search. In standard A ∗ terminology, g ( x ) denotes the accurate and appropriate path from the beginning point to every vertex x, and h ( x ) means the empirical approximate value from vertices n to the target. A ∗ maintains a delicate equilibrium as it progresses by the original position to the objective. Then the parameters such as ϑ, ϖ, x and β represents optimization parameter for optimally reduce the uncertainty problem, such as robot’s linear velocity with angular velocity and robot’s linear acceleration with angular acceleration. These parameters are optimized using A star algorithm. The main loop seeks for the node x to the minimum function f ( x ) = g ( x ) + h ( x ) . The heuristic technique h ( x ) gives A ∗ an estimate of the least expensive path on each nodes n to the destination A. Pseudocode for A Star Algorithm is represented in Algorithm 1.

OPEN IN VIEWER

Algorithm 1

AStar (point cloud data)

3.3. Grid mapping

The Gmapping method is employed as a HSAStar approach in this investigation. It employs the particle filter to acquire lattice representation after laser range sensor data. Using laser scan measurement data z 1 : t = x 1 , … , x t , the joint posterior (4) between map m t and the robot’s state x 1 : t = x 1 , … , x t has been computed from the laser scan z 1 : t = z 1 , … , z t . (7) P ( x 1 : t , m | z 1 : t , u 1 : t − 1 ) Although the EKF and other Kalman filter-based algorithms are good for expressing linearized distributions, the RBPF is a better technique for representing non-gaussian distributions, which is what the test situations require. The core concept of RBPF is to utilize laser sensor data to set the state of hypotheses, where each particle keeps its state. The matching particle is linked to each landmark. The strongest hypotheses are preserved with the provided weights, while the weak ones are eliminated after resampling [25]. Each particle represents a possible robot position, and each state indicates the outcome. (8) P ( x 1 : t , m | z 1 : t , u 0 : t − 1 ) = P ( x 1 : t | z 1 : t , u 0 : t − 1 ) ∗ P ( m | x 1 : t , z 1 : t ) Each particle filter can hold a map, which is chosen based on the likelihood of the inspection in relation a map of its own, as indicated in the probability proportional equation (14). The aim of this work is to find a costless solution to one of SLAM’s most difficult localization problems. This work recommends avoiding data for verification control in identifying posture data from the odometer or other sources by employing Gmapping technique without odometer data. Data from a single laser sensor is used to define pose estimate. The flow of the Gmapping method without control data is shown in Fig. 1. For example, x t denotes state, z t denotes measurement, and m t denotes the derived map. The control data does not appear to be used in Fig. 1. This algorithm’s emulates three steps: guessing the original state, the attitude is derived after the preceding one using the z t − 1 measurement from the initial state guess, the Data-matching method obtains map m t − 1 particles are updated and the last one is particles’ updates are based on z t readings. The particle’s map m t is simplified based on the condition x and the measurement z t . Data-matching is a technique for aligning a scanned map with a built map. The current scan z t aligned with the specified beginning state x t 0 and map m t − 1 using Data-matching [25,51]. Starting from the initial state x t , the Data-matching technique is employed to map m t − 1 (15). The method looks for alignment in only a few locations of state x t . (9) x t i = x t − 1 i + u t − 1 x t i = argmax ( P ( x | m t − 1 1 , z t , x t i ) ) Data-matching is used to pick particles with high hypothesis of state t x (16). As a result of this, the next state is determined as follows: (10) P ( z t | m t − 1 , x j ) → P ( x j , x t − 1 , u t − 1 ) Weights are computed by the normalizer in this phase (11), (11) n ( i ) = ∑ j = 1 k p ( z t | m t − 1 , ( i ) x j ) p ( x j | x t − 1 ( i ) , u t − 1 ) Based on the state x t and measurement z t , the particle’s map m t at task duration t is updated. Algorithm indicates pseudo-code for pathway planning of PMP. To confirm PMP’s outcome, we add the following three selection criteria: ego_Moving Update, go_back, and visited_Location. When approach is processing, short noticeable track is determined to target perspective. If track induces one of three further criteria, resolution is prepared as per that rule. Hyb SLAM A Star algorithm proposed for generating map along with the path where the autonomous vehicle is moved from source to destination. Pseudocode for Hyb SLAM A Star (point cloud data) is represented in Algorithm 2.

OPEN IN VIEWER

Algorithm 2

Hyb SLAM A Star (point cloud data)

ego_Moving Update: A ego_Moving update adjustment function of pc is set based on the discussion of the reward coefficient pc to increase continuous observation performance. Restraint backwards: The go_back of the UGV will unavoidably occur in path planning because of the heuristic technique’s randomness and the change in payout. When the go_back occurs, however, it is difficult to assess if this decision is sensible or not. As a result, the suggested approach allows the UGV to return to its new path. Visited_Location: UGV is said to be in a loop closure once the path budget exceeds payment or when the budget of various positions are equal. To boost monitoring efficiency, when a loop closure occurs, the viewpoint by the greatest payment of altogether neighboring viewpoints at present prespective is picked as the goal points for the following second.

4. System design

Before we can do SLAM research, we need to put up a mobile robot experimental platform. The main control uses ESP8266 Node MCU microcontroller and Raspberry Pi 4 as a Microprocessor to achieve the benefits of low cost, modularity, flexibility, and expandability of mobile robots. To complete the design of the entire hardware system, the robot bottom stage control panel employs the raspberry driver board for Image processing and ESP8266 Node MCU board to connect to a station point to transmit the data from our ESP32 Camera Module and VL53L1X Lidar Sensor and for the robot’s locomotion we use Gear Micro DC Encoder Motor which is connected to a Node MCU Motor Shield and a Step-Down Transformer to regulate the voltage supply [60].

OPEN IN VIEWER

Fig. 4.

Ground vehicle frame design.

4.1. Body of a mobile robot platform

Mobile robot’s structure made up of 2 layer of acrylic plate. It measures 35 centimeters by 20 centimeters by 15 centimeters and weighs about 3 kg. It may make a range of sensors and devices to fulfill the experiment’s necessities. The authentic thing of a ground vehicle is depicted in Fig. 4. The experimentation is done on unknown environmental along the obstacles of static or dynamic to examine the proposed approach proficiency. Every experimental environment set up in office building’s corridor. Here, run global path planner to differentiate variation of orientation path created via global path planner as well as real trajectory derived from the approach. Nevertheless, the reference path does not take part in robot navigation. The original scenario has no constraints, its 2D map created through concurrent localization with mapping to reflect the quality of “unknown.” The actual environment equipped with constraints for navigation. The robot is unaware of the obstacles and can only detect them while it is navigating. These obstacles spatial distribution is set to considerably intercept the robot trajectory towards the goal. A power unit offers the essential energy of controller unit that oversee any actions enabled by the robot. As a robust less cost solution, the Arduino Uno micro controller equipped with appropriate motor shield. A driving unit is necessary to move the robot at X and Y directions. This is accomplished by dual gearbox equipped with two brushed DC motors. Two extra ball caster wheels offered stability when driving. Several sensors are utilized to track the robot progress, also speed encoder mounted on one wheels, an inertial sensor to sense the robot alignment is required. The digital magnetometer presented information regarding current heading of SLAM robot. A sensor unit is required to sense the environment. To attain a proper field of vision, the unit is mounted in rotating servo motor. The environmental conditions based numerous kinds of sensor can be applied. In this case, a low-cost LIDAR device is employed to scale the distance of any obstacles. Also, an ultrasonic range finder is employed, because laser-base sensors could not detect the glass elements. To identify changes in floor level (where there are stairs), an additional device pointing downwards by 45 degrees is suggested.

OPEN IN VIEWER

Fig. 5.

Architecture of a mobile robot’s hardware system.

The mobile robot walks with the help of a four-wheeled walking gadget. A brush motor 6V Gear Micro DC Encoder [53] propels each tyre, allows the vehicle to run at speeds of up to 1 m/s and swivel at a rate of 6.6 rad/s. The robotic tyre speed and displacement are measured using 360 line AB 2 phase continuous rotary encoder. A robot platform control scheme and server make up majority of the robot manipulator hardware system. Figure 5 depicts the structural block diagram. A principal processor, a VL53L1X Lidar Sensor, Inertial Measurement Unit (IMU) unit, a location controller, 22.2V LIPO Lithium Polymer Battery power supply module makes the open source robot platform control system. It’s not only proficient of autonomously directing and placing the mobile robot. Object detection and other aspects, as well as the right to receive server software to send commands to direct the robot’s movement and a range of sensor data sent over WIFI to the server. An ESP8266 Node MCU Development Board with 1GB of RAM and a range of connectors, such as a built-in 10/100 dynamic network adapter, an 802.11n WIFI wireless USB network card, and a server wireless network card, serves as main controller (3D printed using PLA material).

The bottom driver board Node MCU and Raspberry pi 4 is the key component of the position control module. The Node MCU Motor Shield will be the speed controller for 366PRM DC brush motor. The Motor Shield measures the speed at which the object moves and the direction of the robot manure is controlled by movement of side wheels. Determine the robot’s orientation, direction, and acceleration in 3D space, you’re moving. Motor Shield employs PID control to enable the robot to interact with the main controller as well as move precisely. The speed, angle, mileage of robot’s wheels determine its location. The main controller uses the VL53L1X Lidar Sensor laser sensor to capture 2D laser data from the environment, then wirelessly transfers it to the server utilizing ROS distributed processing to make a map of world around the robot.

A 22.2V 2800mAh large rechargeable lithium battery powers the mobile robot platform. Three 18650 lithiumion battery packs make up the battery, which is enough to power the complete system. Lenovo Y480 machines are employed on the server. Ubuntu 16.04 and the Free Software Robotics Operating System are installed in a virtual computer (ROS). The mobile robot platform’s principal controller first shifts the robot’s 2D laser data into 3D space, viz, orientation, velocity, distance, robot’s environment. Map of complete unfamiliar part is regularly generated by managing the robot’s behavior at the host.

4.2. Model of software for mobile robots

The trials are carried out in a two-wheel mobile robot equipped with an Intel Core i7-4500U 1.8 GHz CPU, 8 GB of memory, and a Robot Operating System running in a computer equipped with four Intel Core i7 2.2 GHz processors, eight gigabytes of RAM, and an Intel HD Graphics 6000 graphics card. The ROS system framework builds a platform that links all activities and installs the bulk of the key feature modules and packages [25,51], mostly from the ROS community, after install the complete desktop ROS version. By customizing the robot software system to own robot as well as generating more node and processing packaging, the gained general resources may be simply updated to finish the creation of the robot software system. This study takes extensive usage of ROS distributed processing architecture because map development and vehicle location need a significant count of data processing using a single card type computer is difficult. First, the host and the robot head controller are on the same network and are coupled to the same LAN. On the server, construct Node Managers and keyboard trouble shoots, IMU errors, autonomous vehicle linear and angular velocity adjustments, SLAM nodes in mobile robot host controller, Lidar, Odometer, and Base Controller nodes.

The node manager can then employ a peer-to-peer architecture on the current web to execute TCP/IP message and achieve successful nodes on separate hosts can communicate with each other after all of the nodes have been registered and monitored by the node management. Finally, perform the SLAM job on the server in an unknown area using the 3D visualization application RVIZ. The node topology is considered for robots in the ROS system framework. The processes that the RVIZ tool may debug are mostly handled by the server-side node: To implement SLAM and run at RVIZ, obtained objects such as lasers signals and postures from the robot control system laser sensor and odometer node, respectively. There are options in both horizontal and angular velocity adjustments. To improve the precision of adjustment, change the line tempo and its correction codes.

4.3. Programming for point regulator

As a node of ROS, the point regulator mode analyses the straight speed with angular velocity data provided through the core processing center, to determine parameters that the robot must be able to move in both wheels. The robot follows the trajectory given by the host’s movement instruction and uses laser distance meter to avoid barrier. Finally, robot communicates the mileage node attitude parameters including flow rate, angular momentum, and movement. Figure 6 shows the flowchart for the software.

OPEN IN VIEWER

Fig. 6.

Programming for point regulator.

5. Experimental results and analysis

Following the software requirements for mobile robots, has been built up, the foundation for SLAM research is Robot movement can be controlled remotely using the ROS shared processing system. As a result, we’ll need to use RVIZ and other visualization apparatuses to remotely manage receive comments from the mobile robot at first. After that, SLAM test can utilized to ensure that design fits the need. IMU on trolley should be measured before wheeled operation, to better place the mobile robot operation is controlled. In this study, the GY-85 9-channel IMU is used. IMU inertial sensing sensor featuring triaxial geomagnetometer, triaxial accelerometer, and triaxial gyroscope [60]. The mobile robot’s IMU passes a multiple tests, and the data is analyzed and the coordinates are determined. The x, y, and z systems of mobile robot have been rectified. Way to take the x, y, z coordinate system is constant. To improve the initial motion estimate, IMU’s fusion with a weighted average the initial position estimation and calculation value were used.

Before trolley procedure is managed, IMU on trolley should be measured in order to enhance position the mobile robot. In this study, the GY-85 9-channel IMU is used. Inertial measurement unit (IMU) sensor with axis gyroscope sensor, the triaxial geomagnetometer, triaxial accelerometer, and triaxial gyroscope [53] are all triaxial devices. The mobile robot’s IMU passes a test. A significant number of coordinates have been determined and the data has been calculated. The mobile robot’s x, y, and z systems have been calibrated. The best way to travel the three axis is to use a coordinate system that is consistent. The IMU as well as the IMU’s weighted average fusion is done to enhance the estimation of starting motion by combine value of initial attitude assessment and the value of the attitude calculation.

Figure 7: Top left shows simulated environment. Top Right: Generation of Map with initial guess. Bottom Left: Scan Matching with Map update. Bottom Right: Generated map with obstacle detection.

The SLAM code is injected into the Node MCU module. It is connected directly with Node MCU Motor Shield. The battery, DC motors, and Step down Transformer are all connected to the Motor Shield. The voltage supply required by each item is regulated by the Step down Transformer. The Motor shield aids in the management of the motors. For the Lidar to cover a large angle and project in 180°, it is first positioned on top of the stepper motor. The data from the Lidar is then transferred to the Node MCU, where it is mapped inside RVIZ and sent to SLAM for localization. These data are transferred to our server via Node MCU over WIFI. Based on the Lidar data we receive, the Motor shield supplies us with co-ordinates for robot navigation, and therefore obstacles detection occurs. The Node MCU is attached to the camera module, which captures data as we move and sends it to the Raspberry Pi 4, which has our illness classification module.

OPEN IN VIEWER

Fig. 7.

Autonomous navigation on agriculture field.

The camera as well as ph and humidity sensors, are mounted on a robot arm to provide leverage for capturing images of leaves and testing the soil in the ground. All of these data are submitted to our server, which then generates a report. Figure 7 represents the simulated environment for testing purpose. It simulates initial guess green color assumed as tree, black color assumed as rigid or boundary, gray color treated as agriculture field, white color dots treated as plants in agriculture field. Tree, boundary, plant are consider as obstacle by the propose algorithm. Semantically organized point cloud data is shown in Fig. 8. Left top shows the camera recording and the left bottom shows corresponding Lidar capturing. Right image displays the generated point cloud data.

OPEN IN VIEWER

Fig. 8.

Semantically organized point cloud data.

Most of the autonomous navigation research used image processing based path planning. Figure 9 depicts the Comparison of generated path with SLAM-HSAStar algorithm, LS Track, and EKF Track. Here black color line represents the proposed hyb SLAM-AStar algorithm track and it shows the shortest path compared with existing tracks such as Link state algorithm (LS) Track, and extended Kalman filter (EKF) Track respectively.

OPEN IN VIEWER

Fig. 9.

Comparison of generated path with SLAM, LS track, and EKF track.

OPEN IN VIEWER

Fig. 10.

Pose estimation of autonomous robot.

Pose estimation another important measure for autonomous navigation. Figure 10 shows the pose estimation for different trajectory for unknown environment. Pose estimation influences by the temporal for the optimal route taken by the robot. In the above Fig. 10 X and Y axis represented land coverage in meter, in that black color triangles represent the actual pose estimation in unknown environment. Blue color triangles show the proposed algorithm pose estimation in the unknown environment. Ls track means that the path generated only from the Lidar sensor. Orange color triangle shows that state of art algorithm pose estimation in an unknown environment. Form the graph we infer that the proposed algorithm gives the accurate pose estimation in an unknown environment.

6. Discussion

This study looked into modelling and path of a UGV’s persistent monitoring of an unknown environment. A unique occurrence modelling strategy is planned to discretize path, compute the observing effect, and define the model we are concerned about. Simulation analysis and baseline comparison are used to assess the feasibility and correctness of a path planning method. Simulation demonstrates the method’s validity in the case of weighted visits to certain positions.

The findings of simulation studies show a correlation between system characteristics and algorithm performance, indicating that the actual algorithm performance cannot be anticipated precisely under multiple scenarios. Figure 11 shows the simulated environment of agriculture field. The starting robot’s orientation is depicted in this middle photograph in the simulated environment. Bottom image represents autonomous exploration of robot, in that green color shows the visible objects with in a range. Blue color line interprets the estimated expected path for robot, red color line shows actual navigation of robot and gray color region shows exploration without any obstacle.

OPEN IN VIEWER

Fig. 11.

Simulated environment (top). Initial position of robot (middle). Autonomous exploration of robot (bottom).

The following measures are used to assess the quality of a SLAM algorithm: The number of corners in a map, which determines the precision of a map, the percentage of occupied and free cells, which allows one to identify whether obstacle/object on a map are blurred and whether extra obstacle/object developed owing to an algorithm failure. If there are overlapping obstacle/object or artefacts in an unknown location, this measure can be used to detect them, as can the number of enclosed areas, which expresses the same idea. Table 1 shows the three trials of mapping parameter.

OPEN IN VIEWER

Table 1

Three trials of mapping parameter

Trail	1	2	3
Speed of the robot (m/s)	0.5	0.1	0.12
Map update interval (s)	10	5	3
Particle filter count	30	25	20

OPEN IN VIEWER

Fig. 12.

Procedure of robot experimentation (a)–(e) displays robot location at chronological sort.

As a result, it is difficult to come up with a models, there is a common adaptive criterion of the highest reward coefficient. Precision of map is computed, when the robot has done mapping, the accuracy of the map is calculated. The correctness of map is determined using equation (15), which is depends on the total area of the actual map layout. The x shows the complete range of the robot’s map, meanwhile the y reflects the real map’s overall length.

6.1. Experimental scenario map or scenario for some sample trails

The map or scenario for some sample trails, A star algorithm is upgraded for feasible performance of navigation, then the trained agent is acquired via complex environment having certain trails, trial 1 is given in Fig. 12, trial 2 is given in Fig. 13, trial 3 is given in Fig. 14. The proposed approach handles using the picture including simple changing constraints. Original scenario map fabricated through Simultaneous Localization and Mapping is given in Fig. 13(a), the real environment to navigation is represented in Fig. 13(b). Here, a person randomly walks at scenario who considerably intercepts the robot path twice to target. Figure 14 depicts temporal series of robot navigation snapshots visualized in ROS, viz, related positioning of real robot in experiment. Figure 14(a) portrays initial place, objective and orientation path. Reference path goes straight to goal owing to lack of obstacle in the original map. Figure 14(b)–(g) displays the robot movements averting the moving pedestrian. The robot steers away from this pedestrian when the pedestrian hinders the robot motion (Figs 14(b) to 14(f)). If the person is away from the robot, the robot back to the direction of goal (Figs 14(d), (g)). The real path as robot attains the target is depicted in Fig. 14(e). Here, the real trajectory averts the moving pedestrian, also vary from reference path that exemplifies the proposed approach proficiently navigate the robot in simple dynamic situation.

OPEN IN VIEWER

Fig. 13.

(a): Experiment scenario original map of sample trails 1, (b) real experiment scenario.

OPEN IN VIEWER

Fig. 14.

Procedure of robot experimentation 2. (a)–(h) displays robot location at chronological sort.

6.2. Performance metrics

Here, the performance measures like accuracy, map size, specificity, patch length, map duration and execution time are looked at in order to assess performance.

6.2.1. Accuracy

Accuracy is described as the total number of instances in the data. As a result, equation (12) determines it, (12) Accuracy = ( TP + TN ) ( TP + TN + FP + FN ) where TP denotes the true positive rate, TN indicates the true negative rate, FP represents the false positive and FN indicates the false negative.

6.2.2. Map size

Map size depends on the specific application or system being evaluated. It is estimated using equation (13), (13) Precision of mapsize = map updated by time t map updated by time t + ∇ t ∗ 100

6.2.3. Patch length

Patch length is a term used in environmental monitoring to refer to the size of the area that is being sampled or monitored for a particular environmental parameter. The patch length is usually defined as the distance over which the parameter of interest is assumed to be homogeneous and representative of the larger area being monitored. Patch length is estimated using equation (14), (14) Patch Length = [ ( 2 D ∇ t α ) ] 1 / 2 where, D denotes the diffusion coefficient of the substance being monitored, ∇ t indicates the time interval over which measurements are being taken and α is the desired level of statistical confidence.

6.2.4. Map duration

Map duration is a term used in environmental monitoring to refer to the length of time over which data is collected at each monitoring location to create a spatial map of the environmental parameter of interest. Map duration is estimated using equation (15) as follows, (15) Map Duration = ( ( Z σ ) δ ) 2 where Z is the standard normal variate corresponding to the desired level of statistical confidence, σ is the standard deviation of the environmental parameter being measured, and δ is the desired level of precision (i.e., the maximum acceptable difference between the true value and the estimated value).

6.2.5. Execution time

Execution time is a term used in environmental monitoring to refer to the total amount of time required to complete a monitoring program. The execution time is an important consideration in environmental monitoring, as it affects the overall cost and efficiency of the program.

The execution time can be mathematically defined using the following equation (16), (16) T = n ∗ ( T m + T a ) where T is the execution time (in seconds), n is the number of monitoring locations, T m is the time required to collect data at each monitoring location, and T a is the time required for data analysis.

The map can saved periodically; for the mapping procedure in this study, the SLAM approach is combined with the GMapping algorithm. Three trials with varying robot velocity, map update frequency, and particle filter parameters are established to see how they affect mapping quality. Figure 15 represent the mapping accuracy for different trails. The mapping and navigation process can be observed using the ROS “rviz” interface. Inside the ROS, the multisensory data provided by the RPLIDAR360o laser scanning and two optical progressive encoders were processed. It’s more convenient and quick this way. As it develops, it becomes a method for long-distance navigation and mapping.

OPEN IN VIEWER

Fig. 15.

Accuracy analysis.

Figure 15 shows the accuracy analysis. The proposed hyb SLAM-A Star-APP method provides 22.35%, 29.67%, and 28.67% higher accuracy for trial 1; 32.56%, 29.06% and 29.05% higher accuracy for trial 2; 25.32%, 35.48% and 54.28% higher accuracy for trial 3 than the existing methods such as ACO-APF-APP, APFA-APP, GWO-APP and PSO-APP respectively.

Figure 16 shows the map size analysis. The proposed hyb SLAM-A Star-APP method provides 23.65%, 29.67%, and 32.56% lower map size for trial 1; 36.52%, 22.17% and 34.89% lower map size for trial 2; 31.08%, 29.63% and 32.56% lower map size for trial 3 than the existing methods such as ACO-APF-APP, APFA-APP, GWO-APP and PSO-APP respectively.

OPEN IN VIEWER

Fig. 16.

Map size analysis.

OPEN IN VIEWER

Fig. 17.

Map duration analysis.

OPEN IN VIEWER

Fig. 18.

Patch length analysis.

Figure 17 shows the map duration analysis. The proposed hyb SLAM-A Star-APP method provides 36.29%, 27.68%, and 35.68% lower map duration for trial 1; 21.36%, 29.64% and 29.67% lower map duration for trial 2; 41.32%, 30.27% and 42.35% lower map duration for trial 3 than the existing methods such as ACO-APF-APP, APFA-APP, GWO-APP and PSO-APP respectively.

Figure 18 shows the patch length analysis. The proposed hyb SLAM-A Star-APP method provides 23.65%, 29.68%, and 24.58% lower patch length for map 2; 27.38%, 33.21% and 29.68% lower patch length for map 4; 23.65%, 29.36% and 31.05% lower patch length for map 6 than the existing methods such as ACO-APF-APP, APFA-APP, GWO-APP and PSO-APP respectively.

OPEN IN VIEWER

Fig. 19.

Execution time analysis.

Figure 19 shows the execution time analysis. The proposed hyb SLAM-A Star-APP method provides 32.10%, 29.68%, and 22.47% lower execution time for map 2; 32.15%, 33.21% and 32.10% lower execution time for map 4; 23.10%, 45.36% and 22.10% lower execution time for map 6 than the existing methods, such as ACO-APF-APP, APFA-APP, GWO-APP and PSO-APP respectively.