Intelligent traffic cloud computing system based on ant colony algorithm

Abstract

With the growth of data volume in transportation system, requirements of big data technologies are rapidly increasing. This paper presented an improved ant colony algorithm by using data analysis technologies of cloud computing and data mining. And the influence of different spatio-temporal feature fusion methods on the steering wheel angle value of intelligent vehicles is explored by feature fusion method. Furthermore, time-constrained and space-constrained networks are utilized to extract the key features that affect the steering wheel angle value. Experimental results show that the proposed algorithm improves the efficiency of data processing and information search by 35%, comparing to traditional ant colony algorithm. It is very effective in the shortest path analysis of ITS. Our research shows that the application of real-time information in the logistics distribution system can make the planning process more dynamic and the prediction results closer to reality.

Keywords

Cloud computing data mining ant colony algorithm intelligent transportation system neural network feature fusion

1 Introduction

With the development of science and technology and economy, tremendous pressure has been put on urban transport infrastructure in China because of too many vehicles [1]. For the urban traffic congestion, an intelligent transportation system is proposed. The shortest path algorithm is one of the most critical problems of the intelligent transportation system [2]. Therefore, on the basis of existing traffic resources, it is necessary to control and allocate traffic reasonably and effectively, so as to maximize the effectiveness of traffic. [3, 4]. In 2013, research on the structure and application of business intelligence and data mining in cloud computing was proposed [5]. This shows the extensiveness of cloud computing and data mining in real life. With the popularity and development of computers, cloud computing and data mining have shown great practical value [6]. Many related applications are based on problem-solving methods, such as the optimal path problem in traffic routes [7]. Through data mining technology, the information is deeply analyzed to improve the efficiency of logistics operation and reduce the logistics cost of the entire logistics industry. In the field of traffic flow management and control, this paper explores a new data mining technology to replace the traditional data analysis and interpretation methods of ITS to solve this problem [8, 9].

In a traffic network, there are overlapping bus stations on different routes. If a bus station is regarded as a node, when analyzing the network, it is necessary to abstract the bus stations of different routes, which are similar in space, into a node [10, 11]. This research provides theoretical support for the design and implementation of the shortest path system of the intelligent transportation system. Since then, soft computing research in big data intelligent transportation system has also been put forward [12]. it is more advantageous to use large data and cloud computing analysis technology in the shortest path analysis design. When searching for the best path problem, in the face of the huge number of nodes and complex paths in the search space, an excellent solution is needed [13]. According to the knowledge of graph theory, the speed of search depends on the length of the path experienced in the search process and the number of nodes traversed. It is of far-reaching significance for the long-term development of China’s big cities to give full play to the role of existing transportation facilities, to guide the distribution of traffic flow on the road network and to improve its utilization rate [14].

The emergence of cloud computing provides a new idea for logistics distribution. The concept of logistics cloud is introduced into the distribution link. Under the condition of new network information support, logistics distribution model is constructed, and distribution routes and warehousing are optimized [15]. In this paper, an artificial ant colony algorithm (ACA), which simulates ant behavior, is proposed. This algorithm has the characteristics of distributed computing, positive information feedback and heuristic search. According to the traditional Dijkstra algorithm, the network graph will be more complicated, and there will be more vertices and edges [16]. This will result in a large amount of computation and time overhead, which affects the speed of the algorithm and reduces the efficiency of the algorithm.

This paper uses feature fusion method to construct the network framework of space-time feature fusion, and uses ant colony algorithm to realize the shortest path analysis of transportation in the network environment based on cloud computing and data mining technology. The effect of the designed spatio-temporal feature fusion network is verified.

2 Related work

2.1 Intelligent transportation big data

Applying it to data mining can ensure the availability, certainty and interpretability of data mining results, and can establish a unified model for a system to form a new decision support system. The weight of the path can be negative in the ant colony algorithm, which belongs to the label correction algorithm. Therefore, this paper extracts intersections between roads in the traffic network as the object of separate analysis and then uses intersections as nodes to divide each road into segments, so as to avoid complex and diverse topological relationships between roads.

Table 1
Arc segment table in a data processing

Field name Field length Number of nodes

Arc section ID 6 10

Terminal node ID 3 4

Field name	Field length	Number of nodes
Arc section ID	6	10
Terminal node ID	3	4

In the process of path planning, massive data is processed, so a simple method of graph representation must be found. Node and arc structure is a common data structure for expressing maps. For every node that belongs to a node set, its data items include node numbers and node ordinates. The specific situation is shown in Table 1 and Fig. 1.

Fig.1

Arc segment table in data processing.

The effectiveness of ITS shortest path analysis focuses on the left-turn traffic with compliance rate using video analysis at signalized intersections. The research group took a camera beside the traffic pole of each intersection to collect video data for 2 months. Traffic cameras are installed at specific locations at each intersection. Camera vision only focuses on vehicles entering the left-turn lane, collects video data of road traffic vehicles, and prepares for the analysis of the shortest path planning and design of intelligent transportation. Figure 2 shows various traffic signs and instructions for vehicles on the road.

Fig.2

Road traffic data collection.

2.2 Path planning technology

For the analysis and planning of the shortest path of intelligenttransportation, it is necessary to analyze and discuss the proportion of vehicles entering or leaving the parking lot and the parking area space. One of the common scenarios is reversal of the vehicle, which enters a vertical street and arrives at the parking space, or transports goods on this street, as shown in Fig. 3. In most cases, vehicles can reach their destination by other routes to avoid backing up into the vertical street, but this will involve detours and delays. Therefore, the shortest path planning problem should be fully considered in this process.

Urban road traffic network is mainly composed of two kinds of elements: road section and intersection. The elements of road section include traffic lights, indication signs and prohibition signs, etc. For the time complexity analysis of the improved ant colony algorithm, the worst case happens when each node needs to be relaxed by all other nodes, and then it calls back to the queue to wait for the next scan. This algorithm is not only applicable to vehicle routing planning, but also to various fields, such as robots and other areas requiring a minimum cost. Introducing heuristic information will improve efficiency. In each iteration of the ant colony algorithm, the nodes in the array will be stored in disorder. To select a node with the smallest weight, all nodes must be scanned once. In the case of large amount of data, this is undoubtedly a key factor restricting the speed of the program. When constructing a database, in a multi-layer PCB board, the routing direction of adjacent layer copper lines should be orthogonal in space, and should not be parallel structure to reduce unnecessary interlayer interference. The region of interest (ROI) refers to a rectangular region composed of the initial node and the target node, which obtains the optimal trajectory when the shortest path problem is actually obtained. According to the analysis of the existing intelligent control strategy, the most important two aspects of traffic signal control are reasonable signal phase and reasonable timing of each phase. In Fig. 4, a preliminary design of the shortest path planning is given for intelligent transportation based on cloud computing and data mining. Some of them are analyzed. It can be noted that the D road is the shortest road compared to the other two roads, and the road B is the longest. All other road projects are ranked in part in this design. No project is considered unbeatable. Therefore, D road is the best choice.

Fig.3

Shortest path planning for vehicle reversing.

Fig.4

Shortest path analysis and selection comparison of three roads.

For the shortest path planning of intelligent transportation, it can not only shorten the driving distance of vehicles, and save resources, time and cost, but also reduce the occurrence of road traffic accidents. As shown in Fig. 5, important areas of vehicle and pedestrian accidents are displayed. Accident location varies from environment to environment, mainly concentrated in the intersection area of commercial areas and roads. The findings help government agencies address problems in areas with high accidents. In addition, the research results support traffic planners to formulate the shortest path analysis solutions and accident prevention measures for ITS.

Fig.5

Schematic diagram of road traffic accident-prone areas.

3 Realization of shortest path analysis system for an intelligent transportation system

3.1 Overall design of system

The shortest path analysis system of transportation is realized according to the system design scheme and based on the second development of a geographic information system toolkit and application development kit. Aiming at traffic flow prediction and control in intelligent transportation system, ant colony algorithm is used as the theoretical support of knowledge representation in data mining to carry out data mining and provide decision support for the intelligent transportation system. In application, the online experimental reservation system realizes the sharing of experimental resources through a computer network, which is not limited by time and space. In fact, backward search from the target node to the original node should produce the same results (as long as the original cost of each section is the same). Although the storage space occupied by the improved forward-associated edge storage structure is w+2n, this method can clearly express the relationship between the node and the arc segment. When designing the PCB of the high-frequency circuit, attention should also be paid to the length of the copper wire. The length of the copper wire must not be an integral multiple of the wavelength of the high-frequency signal. The distance between two adjacent nodes is a weight. If the two nodes are not straight lines, the actual length is measured as the weight between the two points. The adjacency matrix is used to store the network topology. As the number of nodes increases, the computational efficiency and storage efficiency decrease, especially in urban bus stops and bus lines.

In the stage of analyzing the current situation of shortest path planning, attention should be paid to road planning, data collection, transportation and activity system. This stage is also related to the establishment of a mathematical model. The system provides an input data model (traffic supply, traffic demand and land use) and estimates the performance indicators of the system. Figure 6 is a flow chart of traffic road planning.

Fig.6

Flow chart of shortest path system for Intelligent Transportation.

Table 2

Analysis results of two clusters

Clusters of cluster	Time(s)	Example
analysis
Cluster1	36	128.13
Cluster2	27	104.11

When dealing with information sets in data mining, the more records in a cluster, the more significant the commonality of records in the cluster is, and the higher the initial credibility is. On the contrary, the results of two cluster analysis of Cluster 1 and Cluster 2 are shown in Table 2 and Fig. 7.

Fig.7

Analysis results of two clusters.

In order to avoid overburdening the observing system, it will also inform the observing system of the path that can be withdrawn from the observation. Requests for travel time changes can be issued at any time, such as continuous intervals. Urban traffic is a complex giant system. Implementing ITS and solving the problem of urban traffic congestion emphasize the coordination of ITS subsystems and the optimal allocation of traffic flow in the whole urban road network.

3.2 Data mining technology based on ant colony algorithms

In order to achieve high availability and high scalability of the system, the system adopts four-tier structure mode. The bottom layer is cloud processing layer, which provides basic support for distributed storage and distributed computing. The algorithm layer is mainly based on ant colony algorithm for data mining. Therefore, our goal is to find the minimum of the following functions under the constraints of the above formula, which is expressed as follows: $S_{j} = \sum_{i = 1}^{N} w_{ij} X_{i}$ (1)

In a linearly separable case, we can get a decision function. Then we extend this idea to the linear nonseparable case. The function obtained is: $F_{r} = (A_{c} + W tan φ) [1 - \frac{K}{iK} (1 - e^{\frac{iL}{k}})]$ (2)

When analyzing data samples with quadratic model E under noisy conditions, a quadratic function and a quadratic function are used to fit the data according to the principle of empirical risk minimization. Its mathematical model formula is: $E_{RME} = \sqrt{\frac{1}{P} \sum_{P = 1}^{P} {(E_{p})}^{2}}$ (3)

In a set of functions W, an optimal function X is obtained to estimate the dependency relationship to minimize the expected risk. Its mathematical expression is: $w_{ij} = w_{ij} + a (\frac{X_{i}}{m} - w_{ij})$ (4)

We discuss two cases in which the system output a takes only two values a=0, l and makes a set of indicative functions (i.e., functions with only two values of 0 or 1). The loss function can be defined as: $a_{k} = {\begin{matrix} 1, s_{k} > s_{j}, \forall_{j}, k \neq j \\ 0, Other \end{matrix}$ (5)

When the joint probability f (x) is known, the learning problem becomes a function to find the minimum classification error. At this point, the loss function can be defined as: $f (x) = sign [ω^{T} x + b]$ (6)

The estimated density function is W(t), then the loss function can be defined as: $w (t) = w_{2} + (w_{1} - w_{2}) \frac{T - t}{T}$ (7)

According to the data mining technology based on ant colony algorithm, the shortest distance from each vertex to the source point should be calculated concretely. At the same time, after finding the shortest path of each vertex, it is necessary to modify the current shortest distance from other points to the source point. Among them, cloud computing-based information network analysis technology is mainly used to assist the setting and pop-up of the point selection window, showing the types of nodes selected before the shortest path analysis and related notes. The development of an intelligent traffic management system based on the demand for information sharing is relatively high, involving the database including detachment vehicle management office database, traffic violation database and grass-roots traffic management system database. In order to avoid overburdening the observing system, it will also inform the observing system of the path that can be withdrawn from the observation. Requests to ask for changes in travel time can be issued at any time, such as a continuous time interval. Urban transportation is a complex and huge system. The implementation of ITS and the solution of urban traffic congestion emphasize the coordinated operation of the various ITS subsystems and the optimal configuration of traffic flow throughout the urban road network. In order to achieve high availability and high scalability of the system, the system uses a four-layer structure mode that the cloud processing layer serves as the underlying support for distributed storage and distributed computing. The algorithm layer is mainly based on the ant colony algorithm.

Class prior probabilities and class conditional probability densities must be known in pattern recognition problems. To this end, according to the idea of the large number theorem in probability theory, the arithmetic mean is used instead of the mathematical expectation in the above formula, so that: $p = {\begin{matrix} k σ^{2} < σ_{d}^{2}, f (P_{g}^{t}) > f_{d} \\ 0 \end{matrix}$ (8)

When analyzing data samples with quadratic model E under noisy conditions, a quadratic function is used to fit the data according to the principle of empirical risk minimization. Its mathematical model formula is: $E_{RME} = \sqrt{\frac{1}{P} \sum_{p = 1}^{P} {(E_{p})}^{2}}$ (9)

There are two kinds of classification problems. For all functions in the set of indicative functions (including those that minimize empirical risk), the relationship between empirical risk and actual risk is satisfied by probability W at least, shown as follows:

$\begin{matrix} W_{j} = 1 + k \sum_{i = 1}^{n} f_{ij} ln f_{ij} / \sum_{j = 1}^{m} \\ (1 + k \sum_{i = 1}^{n} f_{ij} ln f_{ij}) \end{matrix}$ (10)

It can be seen that the actual risk of the data learning machine consists of two parts which are the empirical risk (training error) and the confidence range. The confidence range is related to the d-dimensional h of the learning machine and the number m of training samples. Therefore, it can be expressed simply by the formula: $W_{j} = d_{j} / \sum_{j = 1}^{m} d_{j}$ (11)

Therefore, our goal is to find the minimum of the following functions under the constraints of the above formula, which is expressed as follows: $S_{j} = \sum_{i = 1}^{N} w_{ij} X_{i}$ (13)

If we implement the improved ant colony algorithm in the form of a queue, we can reduce redundancy. A point may be put into the queue again after being out of the queue, that is, after a point has been improved over other points, it may be improved by itself after a period of time. In order to implement the hierarchical search algorithm and simplify the path planning task, the navigation map database needs to be hierarchically constructed. The higher the number of layers in the hierarchical map database, the fewer details of the abstract expression. The optimized ant colony algorithm has a significant reduction in the amount of computation compared to the traditional ant colony algorithm, and the degree of this reduction becomes more apparent as the number of nodes increases or the map is larger and more complex. As long as the logical relationship of the complex system is excavated, the data mining technology based on ant colony algorithm can be used to describe the system in a complete and simplest form. The system searches both sites at the same time and finds the shortest path from the search results, thus greatly shortening the search time. Among them, location operations need accurate street names or landmark building attributes and user-defined location attributes, and other data path optimization decisions need accurate road length, road grade and other data closely related to the actual operation of vehicles. Based on cloud computing and data mining technology, the shortest path analysis technology of intelligent transportation is believed to be able to solve the urban traffic problems.

Fig.8

Different characteristic pooling structures (“C” means stacked convolution layer. Blue, green, yellow and orange boxes represent maximum pooling, time domain convolution, full connection and saftmax layer, respectively.)

4 Intelligent transportation decision making method

4.1 Design of Spatio-temporal Feature Fusion Recursive Neural Network

In deep neural networks, feature fusion method can fuse two or more different features and promote network convergence. Four kinds of pooling operations are shown in Fig. 8. Setting the maximum pooling behind the last convolution layer can help the network learn video-level and fragment-level features. In addition, the latter pooling structure performs worse than other structures, which indicates that it is very important to preserve spatial information in the time dimension of pooling operation. Because of the role of convolution layer and pooling layer, image features reduce in the process of backward transmission, and corresponding features extracted from each layer are also lost layer by layer. In order to avoid the feature loss effect caused by the network backward transmission, the network layer behind the feature pyramid can also receive the features extracted from the front layer, and fuse these features as the input of the layer. This makes the latter sub-network layer supplement the loss feature in forwarding transmission, and the loss feature may contain the edge contour information of the target object. Therefore, the feature pyramid network has a stronger feature extraction ability than the traditional forward network. It can be seen from this that the feature extraction from the front layer of the network and the feature fusion from the back layer in some way can enhance the flow of information in the network and increase the expressive ability of the neural network.

This chapter explores the influence of different spatio-temporal feature fusion methods on the steering wheel angle of network prediction from the perspective of feature fusion. Time-constrained network and space-constrained network extract the key features that affect the steering wheel angle value, and these key features determine the orders of output steering wheel angle value. Therefore, this chapter will design the feature fusion network on the basis of the space-time constrained network, and explore the impact of four feature fusion methods on the angle value of the network prediction steering wheel. In this chapter, four space-time feature fusion methods and the framework of space-time feature fusion network are introduced first, and then experiments are conducted on Udacity dataset to verify the effect of the designed space-time feature fusion network.

4.2 Action layer of spatiotemporal feature fusion recursive neural network

Spatiotemporal feature fusion network consists of convolution layer, LSTM layer, Pooling layer, Merge layer and full connection layer, convolution layer and LSTM layer. Pooling layer is mainly used to reduce the size of feature maps generated by the convolution network and reduce network parameters. Because some elements in the feature graph are generated by repeated computations during convolution operation, a large amount of redundant information is saved in the feature graph. Without pooling, the network needs more parameters to handle these redundant elements. As the number of convolution layers deepens, the amount of computation required by the network increases rapidly. The commonly used pooling methods are Max-Pooling, Average-Pooling and Stochastic-Pooling. The formulas of the first two pooling methods are shown in (14) and (15): $x_{ij}^{l + 1} = max ({x_{si + δ i}^{l}, x_{sj + δ j}^{l}}, 0 \leq δ i, δ j \leq k)$ (14) $x_{ij}^{l + 1} = \frac{1}{k^{2}} \sum_{δ i, δ j} ({x_{si + δ i}^{l}, x_{sj + δ j}^{l}}, 0 \leq δ i, δ j \leq k)$ (15)

Among them, the output values of the rectangular area with length and width of K are pooled, and I and j represent the abscissa and ordinate coordinates of the input pixels, respectively. From the above formula, it can be seen that the maximum pooling is to take the largest pixel in an area as the output, and the average pooling is to take the average value of all the pixels in an area as the output. The calculating formulas of Stochastic-Pooling pooling method are as follows: $p_{i} = \frac{a_{i}}{\sum_{k \in R_{j}} a_{k}}$ (16) $s_{j} = a_{l}, l \sim P (p_{1}, . . ., P_{| R_{j} |})$ (17)

Among them, a_k refers to the pixel value of R_jin the rectangular region, p_i refers to the probability of I position, and s_j refers to the output value of the rectangular region. According to the theory of feature extraction, there are two main reasons for the error of feature extraction, including the increase of variance of estimation caused by the limited size of the neighborhood and the deviation of estimation mean caused by the error of convolution layer parameters. Generally, average pooling can reduce the former error and retain more image background information. The maximum pooling can reduce the latter error and retain more information. Random pooling is between the two.

Full connection layer is a matrix multiplication of the input matrix, which is equivalent to feature space transformation. It aims to extract and integrate useful information. In theory, any non-linear transformation can be simulated by adding the multi-layer full connection of the non-linear mapping of the activation function. Full connection layer can also transform dimension from high dimension to low dimension without losing useful information. The formula of calculating the full connection layer is as follow: $y = Wx + b$ (18)

Where x denotes the input vector or matrix, W denotes the input weight matrix, and b denotes the bias. Formula (18) shows that when the dimension of input and output is relatively high, the size of weight matrix W will become particularly high, which will lead to a sharp increase in network parameters, which is not conducive to the training of network models. The structure of the full connection layer is shown in Fig. 9.

Fig.10

Framework diagram of spatiotemporal feature fusion network.

Fig.11

Framework of the spatio-temporal feature fusion network without spatial location constraints.

4.3 Spatiotemporal feature fusion network framework

Fig.9

Full connection layer structure diagram.

Compared with the spatio-temporal constraint networks, the spatio-temporal feature fusion networks transfer the spatial location features extracted from the last convolution layer to the feature fusion layer, which fuses the incoming spatial location features and temporal context features to generate new fusion features for subsequent sub-networks to learn. The network framework is shown in Fig. 10.

In space-time feature fusion network, both temporal context feature and spatial location feature are time series data, which cannot be fused directly, and need feature pooling. In addition, the activation function, pooling function and dropout layer of spatiotemporal feature fusion network are not shown in Fig. 11.

The space-time feature fusion network designed in this chapter mainly considers the following factors:

The spatial location feature extracted by convolution network can help decision-making network to make decisions. When it is transmitted to the back sub-network in a fast way, it can increase the amount of information in the network.

The learning characteristics of the neural network layer in front of the decision network will be lost in the process of backward transmission. Feature fusion can reuse this part of the information and theoretically promote the learning of the decision network.

Fused features have more information than single temporal context features and spatial location features. They also have the attributes of two features before fusion, and are the more advanced semantic feature information.

By feature fusion, we can increase the connection between the front layer and the back layer of the network, which makes the feature information flow more quickly in the network.

In order to further verify the impact of feature fusion methods on decision networks, we design a spatiotemporal feature fusion network without spatial constraints, as shown in Fig. 11. Compared with the spatio-temporal feature fusion networks, there is no FC3 full connection layer and loss function for decision networks without spatial constraints. There is only one loss function in the whole network. So the network structure is simpler and there are less the network parameters.

5 Experiments and analysis

The experiments on Udacity dataset are divided into two steps: model training and model testing. The feature fusion network with spatial location constraints has better steering wheel angle prediction ability than the feature fusion network without spatial location constraints. Therefore, this section only trains space-time feature fusion network with spatial location constraints, namely, real-time space-time feature addition network, space-time feature subtraction network, space-time feature multiplication network and cascade network of space-time features. Then, the ability of the model to predict the steering wheel angle is tested on the Udacity Intelligent Vehicle Data Set. The experimental steps are described as follows:

Fig.12

Test results of intelligent vehicle and four-feature fusion network.

In the first step, the Udacity smart car data set is divided into training set and test set. We choose the four scenarios of hmb_2, hmb_3, hmb_4 and hmb_5 as a training set, and the hmb_1 scenario as a test set, where hmb_1 to hmb_5 correspond to the first scenario to the fifth scenario respectively. Because we use the original data set to generate some training samples, the training set and the test set also contain the corresponding scenario generation samples. Then, the training samples are used to train the spatiotemporal feature fusion network in turn, and the network parameters are adjusted to obtain the optimal model.

Table 3

Test results of Udacity smart vehicle and four feature fusion networks

Ranking	Ranks	RMSE
1	Komanda	0.049
2	Rambo	0.058
3	Chauffeur	0.059
Network model	Feature addition	0.064
4	Lookma	0.075
5	Rwightman	0.078
6	Epoch	0.08
Network model	Feature subtraction	0.09
7	Bitas	0.096
8	Ai-world	0.104
Network model	Feature cascade	0.106
9	Bauer	0.107
Network model	Feature multiplication	0.121
10	Fsc3	0.123

In order to compare the predictive effect of the four feature fusion networks on steering wheel angle, we will compare the test results of the four networks on the Udacity Smart Vehicle Public Database, and compare the test results with those of other teams in the Udacity Competition Test Set. The scenario of Udacity test suite is very close to that of a public database, which can approximately compare the performance of the algorithm proposed in this chapter with that of other teams. Table 3 shows the results of the top ten teams in Challenge-2 of Udacity Smart Car Competition. The root mean square error of the best results is 0.048, and that of the tenth team is 0.122. Figure 12 is the test results of four feature fusion networks on Udacity public database. The mean square error root of the feature addition fusion method is the smallest, which is only 0.063, showing that the prediction value is the closest to the baseline value. The mean square error of the feature cascade network is 0.106, which is higher than that of the feature addition and subtraction methods. This shows that the vector cascade feature fusion method cannot help the network to predict the steering wheel angle well. The mean square error of the feature multiplication network is 0.121, which is the largest of the four networks, and its deviation between the predicted value and the baseline value is the largest. It can be seen that the ability of feature cascade and feature multiplication to predict steering wheel angle is worse than that of feature addition and feature subtraction.

Fig.13

Prediction curve of steering wheel angle for 4000 frames of Udacity test set based on feature addition network.

Fig.14

Prediction curve of steering wheel angle for 4000 frames of Udacity test set based on feature subtraction network.

In order to evaluate the similarity between the prediction curve and the reference curve subjectively, we selected 4000 images from the Udacity Intelligent Vehicle Database Test Set to test the four feature fusion networks in this section. From Fig. 13 and 14, it can be seen that the four feature fusion networks have greater jitter at t=200 and t=1000, because the scene images at both times are in shadow and the lane lines in the road are almost invisible. The light is relatively dark at night, and the brightness of the collected scene image is low. Once encountering shadows, the scene is almost invisible. This is quite different from the scene of the training sample. However, at these two moments, the deviation between the predicted value and the reference value of the feature addition network is very small unlike the other three networks, which shows that it has strong robustness to abnormal scenarios. The feature subtraction network and the feature multiplication network can fit the datum curve well in normal scenarios, but when t=200 and t=1000, the network prediction value has a large deviation, and it is almost impossible to predict the steering wheel angle normally. Compared with the feature addition network, its robustness is poor. Although the predicted value of feature cascade network deviates little from the reference value at t=1000, it cannot predict the reference value normally at t=200, and the predicted value deviates greatly from the reference value. In summary, the feature addition network can not only predict the steering wheel angle in normal scenarios, but also has strong robustness to abnormal scenarios, and can make safe decisions in abnormal scenarios. It can be inferred that the feature addition network can predict the steering wheel angle better than the other three networks when encountering extreme scenarios.

6 Conclusions

The decision-making behavior of intelligent vehicles depends on the perception module. If the perception module perceives the surrounding environment information incorrectly, it will transmit the wrong perception information to the decision-making module, which has a very high probability of making the decision-making module wrong. On the other hand, the back end of the decision module is connected with the vehicle control module. The decision-making quality of the decision-making module will directly reflect the driving state of the vehicle, such as the vehicle collision caused by extreme output of the decision quantity. Many research institutes at home and abroad have proposed intelligent vehicle system solutions which greatly promote the development of intelligent vehicle technology. The innovation of this paper is that the improved ant colony algorithm is applied to data mining, and the shortest path analysis model of road traffic based on cloud computing and data mining is established, which can be used in traffic signal timing scheme of the plane intersection. This method has achieved good results. Compared with the traditional ant colony algorithm, the search speed has been significantly improved. This experiment is mainly aimed at the static traffic network with positive path weights, which is still applicable to the dynamic traffic network with time-varying weights. Finally, our next research focus will lie on how to effectively use the data mining information technology analysis platform based on ant colony algorithm for real-time road condition prediction and path guidance.

Footnotes

Acknowledgment

This work was supported by the Henan Science and Technology Research Project (No. 182102310025), PhD foundation of Henan Institute of Engineering (No. Dkj2018002) and Open Fund Project of Key Laboratory of Grain Information Processing and Control (No. KFJJ-2016-201).

References

Abduljalil

, Video Capture Service in the Intelligent Transportation System based on Cloud Computing, International Journal of Computer Applications 97(5) (2014), 35–41.

Huang

S.C

, Jiau

M.K

, Lin

C.H

, A Genetic-Algorithm-Based Approach to Solve Carpool Service Problems in Cloud Computing, IEEE Transactions on Intelligent Transportation Systems 16(1) (2015), 352–364.

Darwish

T.S.J

, Bakar

K.A

, Fog Based Intelligent Transportation Big Data Analytics in The Internet of Vehicles Environment: Motivations, Architecture, Challenges, and Critical Issues, IEEE Access 6(99) (2018), 15679–15701.

Mohamed

, The Relation of Artificial Intelligence with Internet Of Things: A survey, Journal of Cybersecurity and Information Management 1(1) (2020), 30–24.

, Shu

, Wu

, et al., A Three-Dimensional Simulation System for Truck Crane Hoisting Based on Cloud Computing, Intelligent Automation & Soft Computing 19(2) (2013), 133–142.

Anisi

M.H

, Abdullah

A.H

, Efficient Data Reporting in Intelligent Transportation Systems, Networks and Spatial Economics 16(2) (2016), 623–642.

N.K

, Ranjan

, Zubair

, et al., Solving the Secondary Structure Matching Problem in Cryo-EM De Novo Modeling Using a Constrained K-Shortest Path Graph Algorithm, IEEE/ACM Trans Comput Biol Bioinform 11(2) (2014), 419–430.

, Wang

, Jiang

, et al., Multi-objective optimization integration of query interfaces for the Deep Web based on attribute constraints, Data & Knowledge Engineering 86(5) (2013), 38–60.

Mohamed

, A comparative study on Internet of Things (IoT): Frameworks, Tools, Applications and Future directions, Journal of Intelligent Systems and Internet of Things 1(1) (2020), 13–39.

10.

Yuan

N.J

, Zheng

, Zhang

, et al., T-Finder: A Recommender System for Finding Passengers and Vacant Taxis, IEEE Transactions on Knowledge & Data Engineering 25(10) (2013), 2390–2403.

11.

Singh

P.K

, Plithogenic set for multi-variable data analysis, International Journal of Neutrosophic Science 1(2) (2020), 81–89.

12.

Shen

, Zhan

Z.H

, Chen

W.N

, et al., Bi-Velocity Discrete Particle Swarm Optimization and Its Application to Multicast Routing Problem in Communication Networks, IEEE Transactions on Industrial Electronics 61(12) (2014), 7141–7151.

13.

Francesquini

, Márcio

, Penna

, et al., On the Energy Efficiency and Performance of Irregular Application Executions on Multicore, NUMA and Manycore Platforms, Journal of Parallel & Distributed Computing 76(4) (2015), 32–48.

14.

Orgaz

G.B

, Barrero

D.F

, R-Moreno

M.D

, et al., Acquisition of business intelligence from human experience in route planning, Enterprise Information Systems 9(3) (2015), 303–323.

15.

Chmiel

, Dańda

, Dziech

, et al., INSIGMA: an intelligent transportation system for urban mobility enhancement, Multimedia Tools and Applications 75(17) (2016), 10529–10560.

16.

Bedogni

, Bononi

, Di Felice

, et al., A Route Planner Service with Recharging Reservation: Electric Itinerary with a Click, IEEE Intelligent Transportation Systems Magazine 8(3) (2016), 75–84.