Traffic light control method based on multi-channel vehicle detection and stereo feature annotation

Abstract

The present invention discloses a traffic light control method based on multi-channel vehicle detection and three-dimensional feature annotation, belonging to the field of vehicle detection and traffic light control. The method includes using a vehicle detection model to collect vehicle information at intersections; introducing a dual-dimensional temporal and spatial road congestion importance coefficient at important road sections and intersections to form a three-dimensional feature annotation of road network traffic flow; establishing a traffic light control system based on a fuzzy neural network, using the obtained regional road network traffic state congestion index as input to the fuzzy neural network to obtain predicted traffic light control information. The traffic light control method based on multi-channel vehicle detection and three-dimensional feature annotation provided by the present invention considers the factors affecting traffic light control such as multiple intersections, multiple lanes, and the temporal and spatial distribution of vehicles, performs fuzzy information processing through neural network technology, and automatically extracts fuzzy rules and generates fuzzy membership functions, providing a visualization method for observing and studying the evolution of network traffic states.

Description

Technical Field

[0001] The present invention belongs to the field of vehicle detection and traffic light control, and in particular, relates to a traffic light control method based on multi-channel vehicle detection and three-dimensional feature labeling. Background Art

[0002] With the continuous growth of the number of motor vehicles, more and more cities around the world are experiencing traffic overload, which is particularly evident during peak hours in the morning and evening. Currently, there are three main approaches to urban traffic light control: (1) timing control system, in which each traffic light control device has set the order and duration of each phase execution, without considering the dynamic changes in traffic flow; (2) vehicle-driven signal control, which dynamically determines the increase or decrease in the execution duration of a phase through the sensing equipment installed at the intersection. Compared with the timing control system, it already takes into account the dynamic flow of vehicles; (3) traffic light adaptive control, which dynamically and adaptively manages and allocates the execution phase and duration based on the traffic status of vehicles at the current road network intersection. Incorporating machine learning and artificial intelligence technologies into adaptive traffic light control systems can further improve the effectiveness of the system. The fixed cycle duration, fixed phase sequence, and fixed phase ratio in traditional signal light control methods do not have the ability to respond to changes in traffic conditions, often resulting in wasted green light time and low traffic efficiency. In addition, some intelligent traffic light control systems have not conducted in-depth mining and analysis of traffic congestion information, and are unable to meet the control system's requirements for low latency and high reliability. When considering factors for judging the level of traffic status, most of them stay at the congestion characteristics of vehicle queues on two-dimensional plane roads, focusing on the research and analysis of frequently congested sections in space, while ignoring the impact of frequently congested sections in the time dimension on traffic light control. The summary and analysis of information on the road are characterized by being single and flat.

[0003] Researching vehicle detection models for urban arterial roads and accurately performing vehicle detection tasks in real time will facilitate traffic signal optimization and coordination under vehicle-road-cloud collaboration, thereby improving the overall operational efficiency of the road network. In recent years, deep learning, represented by convolutional neural networks, has gradually become a popular research method in the field of vehicle detection. Existing research includes methods such as the candidate box convolutional neural network (RCNN), Fast RCNN, and Faster RCNN; and regression-based methods such as YOLO (you-only-look-once) and ResNet50. The candidate box detection scheme is a two-stage method that achieves high detection accuracy but also requires high resources. The regression detection scheme is a single-stage method that is fast but suffers from issues such as insufficient feature fusion and reduced robustness to continuous operation. Furthermore, existing detection frameworks typically first encode the input image into a low-resolution representation and then recover a high-resolution representation from the encoded low-resolution representation. However, given the high-speed mobility of vehicles on road networks, high-resolution representation is crucial for position-sensitive vehicle detection. The high-resolution network (HRNet) adopts a parallel branch structure, maintains high-resolution representation during the learning process, and supports multi-scale fusion between features of different resolutions, which is conducive to extracting target key point features and realizing small target recognition. Summary of the Invention

[0004] In response to the above-mentioned problems, the present invention provides a traffic light control method based on multi-channel vehicle detection and three-dimensional feature labeling. By considering the influencing factors of multiple intersections, multiple lanes, and vehicle spatiotemporal distribution on traffic light control, neural network technology is used for fuzzy information processing. The fuzzy neural network adaptive signal timing dynamically adjusts the signal timing strategy according to the real-time traffic conditions of the signal intersection, in order to minimize the traffic delays at the signal intersection while meeting the vehicle traffic needs, and ultimately achieve the goal of minimizing the entire social travel cost within the signal intersection.

[0005] The technical solutions adopted by the present invention to solve the above technical problems are as follows:

[0006] The traffic light control method based on multi-channel vehicle detection and three-dimensional feature annotation includes the following steps:

[0007] Step a: Using a vehicle detection model to perform feature detection and collect vehicle information at the intersection;

[0008] Step b: analyzing the vehicle congestion status on the three-dimensional plane road based on the intersection vehicle information collected in step a, adding the importance weight coefficients of each road section and intersection in the road network, obtaining the regional road network traffic status congestion index, and forming a three-dimensional feature annotation of the road network traffic flow;

[0009] Step c: Establish a traffic light control system based on a fuzzy neural network. Use the regional road network traffic congestion index obtained in step b as the input of the fuzzy neural network. Calculate the corresponding relationship between the regional road network traffic congestion status and the traffic light control results over a period of time, obtain predicted traffic light control information, and realize global adaptive signal control of multiple sections.

[0010] Furthermore, the vehicle detection model is a vehicle detection model that integrates a multi-channel high-resolution network with an attention mechanism. The model is divided into a road network perception layer, a training network layer, and a feature output layer. The road network perception layer uses a combination of cameras and radar sensors to operate around the clock in zero light and various weather conditions, realizing vehicle detection, lane line recognition, vehicle distribution, and time recording feature detection. The collected road information is passed to the feature output layer, and digital information of each monitored intersection is extracted as input to the training network layer.

[0011] The feature fusion method of integrating the attention mechanism in the high-resolution network is to use the feature connection weight and weight threshold at the network feature fusion connection to perform multi-resolution feature fusion on multiple adjacent network paths in the parallel network structure to obtain a high-resolution parallel network module, where the weight threshold is a hyperparameter and the feature connection weight is a learnable weight parameter; during the network model training process, the feature connection weight is adaptively learned based on the loss of reverse transmission to express the degree of influence of the current feature map on the network.

[0012] Furthermore, a feature fusion method of skip connection is introduced for the high-resolution network, and the skip connection is used to perform cross-layer feature fusion on the feature maps at symmetrical positions in each network path in the parallel network structure.

[0013] Furthermore, the multi-resolution feature fusion of the feature map in the high-resolution parallel network module is specifically performed as follows: the feature map of the q+1th convolution layer in the p+1 network path in the high-resolution parallel network module is Feature (p+1)(q+1) , where p = 1, 2, ..., R and q = 1, 2, ..., Q p , R is the total number of network paths in the high-resolution cascade network module, Q p is the total number of convolutional layers in the p-th network path, then Feature (p+1)(q+1) The steps of multi-resolution feature fusion are as follows:

[0014] Step a1: Get feature map pq 、Feature p(q+1) and Feature p(q+2) , and obtain the feature connection weight W corresponding to the feature map pq 、W p(q+1) and Wp(q+2) ;

[0015] Among them, Feature pq Represents the feature map of the qth convolutional layer in the pth network path in the high-resolution parallel network module, Feature p(q+1) Represents the feature map of the q+1th convolutional layer in the pth network path, Feature p(q+2) represents the feature map of the q+2th convolutional layer in the pth network path, W pq Represents Feature pq The corresponding feature connection weight, W p(q+1) Represents Feature p(q+1) The corresponding feature connection weight, W p(q+2) Represents Feature p(q+2) The corresponding feature connection weights;

[0016] Step a2: Connect the features with weights W in sequence pq 、W p(q+1) 、W p(q+2) Compared with the weight threshold K, the connection activity parameter is obtained; the feature connection weight W pq Connection activity parameter For example:

[0017]

[0018] Multi-resolution feature fusion is performed based on the connection activity parameters, and the following is obtained:

[0019]

[0020] When the connection weight of a feature is lower than the weight threshold, the feature map corresponding to the feature connection weight will be inactivated, and there is no need to fusion the inactivated feature map in the subsequent multi-resolution feature fusion.

[0021] Furthermore, the feature extraction part of the vehicle detection model uses cross-stage residual blocks to replace the traditional residual blocks in the feature extraction network. The feature fusion part introduces spatial pyramid pooling to fuse local area features with global area features, fully integrating and utilizing the feature information extracted from different deep network layers.

[0022] After image preprocessing and lane line detection on the input image, a median-based background extraction method is used to subtract each frame image from the calculated background image, and the subtracted image is then binarized. A background difference algorithm is then used on the extracted background image to extract the video foreground vehicle target. After extracting the video foreground vehicle target, the image is further segmented and processed into a single lane image, and then feature extraction is performed on the single lane image using projection features. After extracting the single lane features, the single lane feature vector is classified using the K-nearest neighbor algorithm to obtain the final detection result, which provides a decision-making basis for the next step of calculating the intelligent traffic light control solution.

[0023] The expression for the binarization processing of the image is as follows:

[0024]

[0025] Where x, y are the coordinate points of the corresponding image, B(x, y) is the background frame obtained by background extraction, and I i (x, y) represents the i-th frame image in the image sequence, D i (x,y) represents the difference between the point (x,y) on the i-th frame. If the difference between the two images at this point is greater than the threshold T, the result is 1; if the difference is less than the threshold T, the result is 0.

[0026] Furthermore, the process of adding the importance weight coefficients of each road section and intersection in the road network to obtain the regional road network traffic congestion index in step b specifically includes:

[0027] The road network is represented by graph theory, which abstracts the road network into a weighted directed graph, abstracts the intersections into nodes in the road network, and abstracts the road segments between adjacent nodes into edges. The regional road network model composed of intersections and road segments can be described as: G = {I, L}; where I represents the set of intersections in the regional road network, I = {I1, I2, ... I n}, I n Indicates the nth intersection in the regional road network intersection set, L = {l ij |l ij =(I i ,I j ); i, j = 1, 2, ..., n, i ≠ j} represents the set of road segments between adjacent nodes; l ij Indicates that there is a road from intersection I i To intersection I j The road section has a length of l ij ;

[0028] Step b1: First, establish the traffic state model of the road section and intersection, and then calculate and integrate it into the traffic state matrix of the entire regional road network;

[0029] The road section traffic state model is as follows:

[0030]

[0031] in, For section L ij Traffic status value, For section L ij The maximum speed limit, v ij For section L ij The average speed, q ij is the measured flow rate, C ij is the traffic capacity, α and β are model parameters, α + β = 1 and α, β∈[0,1]; when the actual traffic volume of the road section is small, the interference between vehicles is small, the running speed is high, the traffic state is good, and the road section state value is also smaller, and vice versa;

[0032] The intersection traffic state model is as follows:

[0033]

[0034] Among them, P i I is the traffic state value of intersection i; saturation R is the ratio of the actual traffic volume to the traffic capacity of the intersection, and its formula is: Where R is the intersection entrance saturation, Q and C are the actual traffic flow and traffic capacity of the entrance, S is the signal phase saturation flow, c is the signal cycle length, and g is the effective green light time; R i is the maximum saturation of the entrance road of the key section of the intersection; m i is the number of key sections at intersection i;

[0035] According to the regional road network topology, the node adjacency matrix A(G) that represents the physical connection relationship of the actual road network can be obtained. mn , the elements are expressed by the following formula:

[0036]

[0037] Among them, a ij =1 indicates that there are actual passable sections at intersections i and j, otherwise they do not exist or are prohibited;

[0038] By collecting traffic parameters on different road sections, the traffic flow of different road sections and intersections in [t k ,t k+1 ] time period; Based on the adjacency matrix and the traffic status of the road section, the road network adjacency traffic status matrix is constructed, which is recorded as A(P)={p ij (t)}, where p ij (t) can be expressed by the following formula:

[0039]

[0040] where p ij (t) represents the traffic status index of different road sections in the road network at time t; if p ij (t) = ∞ means the road section does not exist or is prohibited;

[0041] Step b2: Considering the traffic capacity per unit length, the spatial importance coefficient is used to reflect the spatial distribution of frequently congested sections and intersections in the road network. The spatial importance coefficients of the sections and intersections are as follows:

[0042]

[0043]

[0044] in, and is the key coefficient, and its value range is [0,1]; m i The road section governed by the intersection;

[0045] The congestion time percentage is used to reflect the frequency and degree of traffic network congestion. The time importance coefficient is defined as the frequency coefficient of a certain degree of congestion in a certain time period of a day in the urban road network. The standards for judging congestion in different time periods are not uniform, so different traffic congestion thresholds γ need to be selected according to the actual situation of the city. If the congestion level of a road section is f t (l ij ) is greater than the traffic congestion threshold, it means it is in a congested state, otherwise it is in a smooth state. Each road section should meet the following conditions:

[0046]

[0047]

[0048] Among them, f t (l ij ) is used to determine the road section l in the tth time period within the statistical days ij The congestion state of Gra t (l ij ) represents the congestion level of a single road section in the tth time period within the statistical days, F ij (t) represents the road segment l ij The total number of frequencies with a congested state, N is the total number of time periods;

[0049] The time importance coefficient is expressed by the time period congestion duration ratio, and the road section time importance coefficient is expressed by the time period congestion duration ratio. and intersection time importance coefficient as follows:

[0050]

[0051]

[0052] Where F(t) is the sum of the frequencies of all road sections being congested;

[0053] The influence of the time importance coefficient and space importance coefficient of the above sections and intersections on the regional road network traffic status is normalized and sorted out to obtain the road network traffic status weight model; section weight model and intersection weight model The formula is as follows:

[0054]

[0055]

[0056] Through the analysis of the spatiotemporal importance weight model of road sections and intersections in the regional road network, the road section weight coefficient matrix W L (t) and the intersection weight vector W I (t) are as follows:

[0057]

[0058] W I (t)=(w1 w2…w n )

[0059] Step b3: Based on the traffic state matrix of the regional road network as a whole, the spatiotemporal importance weights of road sections and intersections are considered. The regional road network state model is as follows:

[0060]

[0061] Among them, P(G) is the regional road network traffic state matrix; the number PI ij (t) is the traffic status of the road network unit. When i≠j, PI ij (t) is the traffic status index of the road section. When i=j, PI ij (t) represents the state index of the intersection. When PI ij When (t)=∞, it means that there is no L ij Road section.PI ij The (t) value reflects the congestion level of a road section or intersection, with larger values indicating more congestion;

[0062] Calculate the value of the traffic congestion index PI of the regional road network as a whole:

[0063]

[0064] Among them, PI is the regional road network traffic congestion index, which is used to indicate the degree of traffic congestion in a sub-region of the road network. Analysis shows that the larger the PI value, the more congested the road network; the smaller the PI value, the smoother the traffic in the entire road network. λ is a model parameter with a value of [0,1].

[0065] Furthermore, a self-organizing fuzzy neural network model is obtained by adding a sequence identifier and a sequence locator module to the fuzzy neural network structure, wherein the sequence identifier module is used to receive the initial sample of the sequence to identify the sequence pattern in different learning sequences, and the sequence locator is used to locate the current state in the sequence and use the two parts as inputs of the fuzzy rule to calculate the next sample of the identified sequence; the self-organizing fuzzy neural network model includes a discrimination layer, a sequence fuzzy layer, a storage layer, a sample fuzzy layer, a fuzzy rule layer, a normalization layer and an output layer, and switches S1, S2 and S3 are set to manage the model. The specific functions of each structural layer are as follows:

[0066] Identification layer: used to calculate the identification number of each sequence. This layer accepts T sequence samples of the traffic state sequence. When x(t) is the input at the tth moment, the output of the pth neuron is expressed as:

[0067]

[0068] Among them, after T samples, the output of the p-th neuron is

[0069] Sequential fuzzy layer: It is used to receive the vector containing all the output information of the discriminator layer and define multivariate fuzzy sets for different identification numbers in the discriminator layer. The output of the pth neuron in the sequential fuzzy layer is expressed as:

[0070]

[0071] Among them, c1 p is the vector of the center of the pth fuzzy set, ∑1 is the diagonal matrix containing the width of the fuzzy set for each input dimension;

[0072] Storage layer: Each neuron in this layer is a discrete low-pass filter with a different time constant, λ p =p / (p+1), the output is a vector containing the current sample and d previous samples of the sequence, the formula is as follows:

[0073]

[0074] Sample fuzzy layer: The principle is similar to the sequence fuzzy layer. The neuron output formula is as follows:

[0075]

[0076] Fuzzy rule layer: Each neuron in this layer represents a fuzzy rule. The T-Norm operator is applied to calculate the membership based on the output of the sample fuzzy layer. The membership function of the pth fuzzy rule is defined as:

[0077]

[0078] Among them, μ p It is the membership degree of the current sequence to the pth fuzzy rule and the current position of its corresponding fuzzy rule.

[0079] Normalization layer: Normalize the calculated membership. The output of the p-th neuron in this layer is as follows:

[0080]

[0081] Among them, φ p is the normalized membership of the current sequence to the pth fuzzy rule and the current position of its corresponding fuzzy rule, and R is the number of fuzzy rules.

[0082] Output layer: The output of this layer is the inner product of the input vector and the weight vector W. The weight vector serves as the consequent parameter. The final output of the network is as follows:

[0083] o (7) =y(t)=x(t+1)=W T o (6)

[0084] During the learning process, switch S1 is closed and S2 is open. For the first T sequence samples, switch S3 is closed, and the model input is the original sequence sample x(t). After receiving T samples, switches S1 and S3 are opened, and S2 is closed. The model uses the output of the storage layer as the first fuzzy set center of the fuzzy set layer, locates the current position of the sequence based on global information, and combines fuzzy rules to obtain the predicted sample.

[0085] The regional road network traffic status congestion index obtained in step b is serialized based on the time parameter and used as the input of the fuzzy neural network model. The sequence law of the specific time period is learned according to the time parameter, and the sequence law is applied to the fuzzy neural network with adaptive function to obtain the traffic congestion status prediction value of the corresponding road section, intersection or regional road section. Based on this, the predicted traffic light control information is obtained for the traffic congestion status prediction value obtained above.

[0086] Furthermore, the fuzzy control rules of the self-organizing fuzzy neural network model are set as follows: the fuzzy control rules of the self-organizing fuzzy neural network model are set as follows: "traffic flow three-dimensional characteristic state information" and "spatiotemporal importance level information" are used as inputs of the traffic light control system, and "traffic light control" is used as output of the traffic light control system; the fuzzy three-dimensional characteristic state information of the traffic flow includes four states of "very smooth, smooth, crowded, and very crowded", and the spatiotemporal importance level information includes four states of "very low, low, high, and very high", which together constitute a total of 16 input states. Based on the adaptive signal control system, the corresponding signal control rules are realized through the fuzzy neural network control algorithm, and traffic light control of five states of "very short, short, medium, long, and very long" can be obtained.

[0087] Furthermore, the fuzzy membership function adopts a triangular membership function, and its function form is as follows:

[0088]

[0089] The specific form and distribution are determined by the parameters [a, b, c], requiring a≤b≤c, and the parameter values are given based on expert experience scores.

[0090] The technical solution of the present invention can produce the following technical effects:

[0091] 1. The traffic light control method based on multi-channel vehicle detection and 3D feature annotation proposed in this paper considers the influencing factors of multiple intersections, multiple lanes, and the temporal and spatial distribution of vehicles on traffic light control. It uses neural network technology to process fuzzy information, automatically extract fuzzy rules, and generate fuzzy membership functions. Based on this, two modules, a sequence identifier and a sequence locator, are added to this method. A self-organizing and learning fuzzy neural network model is proposed to achieve the learning and prediction of habitual task sequences. This provides a visualization method for observing and studying the evolution of network traffic status, and can improve the reliability and autonomous learning adaptability of the traffic light control system.

[0092] 2. To address the imbalance and flatness of the model, this invention combines macroscopic road network traffic status evaluation with microscopic traffic parameters to define a new road network traffic status coefficient. This new coefficient takes into account the spatiotemporal congestion status of road sections and intersections, defining spatiotemporal importance weight coefficients. Based on this, a three-dimensional regional traffic status model is established to analyze and evaluate the traffic status of regional road networks with three-dimensional feature annotation.

[0093] 3. The multi-channel high-resolution network structure designed by this invention studies the centralized training, correction and verification methods of distributed videos from multiple monitoring points. It also integrates the feature fusion method of the attention mechanism and the skip connection operation. It achieves efficient fusion of multi-scale features through self-learning of feature connection weights. This enables the network to quickly integrate contextual information during forward propagation and perform lossless propagation in the loss direction, which can accelerate network convergence and achieve deep fusion of multi-scale features.

[0094] 4. In the feature extraction stage, the present invention uses cross-stage residual blocks to replace the traditional residual blocks in the feature extraction network, and introduces spatial pyramid pooling in the feature fusion part to strengthen the full integration of local and regional features at different depth network layers, further improving the lane line detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 The vehicle detection model that integrates the multi-channel high-resolution Wangluo and attention mechanism described in the present invention;

[0096] Figure 2 Schematic diagram of the feature fusion method of the present invention integrating the attention mechanism;

[0097] Figure 3 The lane line recognition optimization processing framework described in the embodiment of the present invention;

[0098] Figure 4 This is the self-organizing and learning fuzzy neural network model described in the embodiment of the present invention. DETAILED DESCRIPTION

[0099] To make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them.

[0100] The traffic light control method based on multi-channel vehicle detection and three-dimensional feature annotation in this embodiment includes the following steps:

[0101] Step 1: Implement feature detection through vehicle detection model

[0102] In this embodiment, in order to improve road traffic monitoring and ensure the safe, reliable and stable operation of the intelligent connected transportation system, the vehicle detection model is a vehicle detection model that integrates a multi-channel high-resolution network and an attention mechanism, such as Figure 1As shown in the figure, the model is divided into a road network perception layer, a training network layer and a feature output layer, wherein the road network perception layer adopts a joint configuration of cameras and radar sensors to ensure all-weather operation under zero light and various weather conditions, and realizes feature detection functions such as vehicle detection, lane line recognition, vehicle distribution and time recording; and transmits the road information collected by the road network perception layer to the feature output layer, and extracts the digital information of each monitored intersection as the input of the training network layer; the vehicle detection model adopts a multi-channel high-resolution network structure design to retain high-resolution features and ensure recognition accuracy; and introduces a feature fusion method of skip connection for the high-resolution network, and uses skip connection to perform cross-layer feature fusion on the feature maps of symmetrical positions in each network path in the parallel network structure.

[0103] The feature fusion method of integrating the attention mechanism in the high-resolution network is to use the feature connection weight and weight threshold at the network feature fusion connection to perform multi-resolution feature fusion on multiple adjacent network paths in the parallel network structure to obtain a high-resolution parallel network module, where the weight threshold is a hyperparameter and the feature connection weight is a learnable weight parameter;

[0104] Feature fusion methods that integrate attention mechanisms such as Figure 2 As shown in the figure, it uses feature connection weights and weight thresholds at the network feature fusion connection to perform multi-resolution feature fusion on multiple adjacent network paths in the parallel network structure, obtaining a high-resolution parallel network module. The weight threshold is a hyperparameter, and the feature connection weight is a learnable weight parameter. During network model training, the feature connection weights can be adaptively learned based on the backpropagation loss to express the influence (i.e., importance) of the current feature map on the network, thereby outputting more informative and robust multi-scale features.

[0105] Suppose the feature map of the q+1th convolutional layer in the p+1 network path in the high-resolution parallel network module is Feature (p+1)(q+1) , where p = 1, 2, ..., R and q = 1, 2, ..., Q p , R is the total number of network paths in the high-resolution cascade network module, Q p is the total number of convolutional layers in the p-th network path, then Feature (p+1)(q+1) The steps of multi-resolution feature fusion are as follows:

[0106] 1) Get the feature map pq 、Feature p(q+1) and Feature p(q+2) , and obtain the feature connection weight W corresponding to the feature map pq 、W p(q+1) and W p(q+2) Among them, Featurepq Represents the feature map of the qth convolutional layer in the pth network path in the high-resolution parallel network module, Feature p(q+1) Represents the feature map of the q+1th convolutional layer in the pth network path, Feature p(q+2) Represents the feature map W of the q+2th convolutional layer in the pth network path pq Represents Feature pq The corresponding feature connection weight, W p(q+1) Represents Feature p(q+1) The corresponding feature connection weight, W p(q+2) Represents Feature p(q+2) The corresponding feature connection weights.

[0107] 2) Connect the feature weights W in sequence pq 、W p(q+1) 、W p(q+2) Compare with the weight threshold K to obtain the connection activity parameter. pq Connection activity parameter For example:

[0108]

[0109] Multi-resolution feature fusion is performed based on the connection activity parameters, and the following is obtained:

[0110]

[0111] When a feature connection weight falls below a weight threshold, the feature map corresponding to that weight is deactivated. Subsequent multi-resolution feature fusion does not require the deactivated feature map. Therefore, by combining feature connection weights and weight thresholds, we can achieve a sparse network and reduce overfitting.

[0112] The lane detection process can be divided into four steps: (1) image preprocessing; (2) feature extraction; (3) feature fusion; and (4) target recognition. To further improve the lane detection performance, the feature extraction part of the vehicle detection model uses cross-stage residual blocks to replace the traditional residual blocks in the feature extraction network. The feature fusion part introduces spatial pyramid pooling to fuse local region features with global region features, fully integrating and utilizing the feature information extracted from different depth network layers.

[0113] like Figure 3The figure shows the lane line recognition optimization processing framework. First, a training set is selected from the dataset and the lane line data in the training set is preprocessed. After preprocessing, irrelevant interference factors in the image can be eliminated, and the processed image in a unified form is used as the input of the entire training network. After image preprocessing and lane line detection operations are performed on the input image, a median-based background extraction method is used to subtract each frame image from the calculated background image, and the subtracted image is binarized. Then, a background difference algorithm is used on the extracted background image to extract the video foreground vehicle target. After extracting the video foreground vehicle target, the image is further segmented and processed into a single lane image, and then the projection feature is used to extract the features of the single lane image. After extracting the single lane feature, since there are five types of traffic flow conditions in each direction in the intelligent traffic light control scheme: very few (VS), few (S), normal (N), many (R) and very many (VR), this requires that the results of the video-based road traffic flow detection algorithm also need to be divided into five categories. Therefore, the K-nearest neighbor algorithm is used to classify the single lane feature vector to obtain the final detection result, thereby providing a decision basis for the next step of calculating the intelligent traffic light control scheme. The expression for the binarization of the image is as follows:

[0114]

[0115] Where x, y are the coordinate points of the corresponding image, B(x, y) is the background frame obtained by background extraction, and I i (x, y) represents the i-th frame image in the image sequence, D i (x,y) represents the difference between the point (x,y) on the i-th frame. If the difference between the two images at this point is greater than the threshold T, the result is 1; if the difference is less than the threshold T, the result is 0.

[0116] The extracted features are fed into the improved pyramid network for feature fusion. The fused feature maps are then fed into the YOLO layer for training. After reaching the training threshold, the iterations are terminated and the final lane detection model is generated. The test set images are fed into the same network, and the trained lane detection model is used to perform lane detection on the test set images, outputting the lane detection results.

[0117] Step 2: The intersection vehicle information collected in step 1 is used as the traffic congestion status data of each regional road network. The vehicle congestion status on the three-dimensional plane road is analyzed, and the importance weight coefficient of each road section and intersection in the road network is added to obtain the regional road network traffic status congestion index, forming a three-dimensional feature annotation of the road network traffic flow.

[0118] The lane detection model in step 1 is used to delineate the borders of each lane. The distribution of vehicles in each lane is identified by combining radar sensors and video detection technology. The importance weight coefficients of each road section and intersection in the road network are added to realize the presentation of three-dimensional feature annotation of road network traffic flow. The process of adding the importance weight coefficients of each road section and intersection in the road network to obtain the regional road network traffic status congestion index and forming the three-dimensional feature annotation of road network traffic flow specifically includes:

[0119] This embodiment uses graph theory to represent the road network, abstracting the road network into a weighted directed graph, abstracting the intersections into nodes in the road network, and abstracting the road segments between adjacent nodes into edges. The regional road network model composed of intersections and road segments can be described as: G = {I, L}; where I represents the set of intersections in the regional road network, I = {I1, I2, ... I n}, I n represents the nth intersection in the regional road network intersection set, L={l ij |l ij =(I i ,I j ); i, j = 1, 2, ..., n, i ≠ j} represents the set of road segments between adjacent nodes; l ij Indicates that there is a road from intersection I i To intersection I j The road section has a length of l ij ;

[0120] Step 21: First, establish a traffic congestion state model for road sections and intersections, and then calculate and integrate it into an overall traffic state matrix for the regional road network;

[0121] The traffic congestion state model of the road section is as follows:

[0122]

[0123] in, For section L ij Traffic status value, For section L ij The maximum speed limit, v ij For section L ij The average speed, q ij is the measured flow rate, C ij is the traffic capacity, α and β are model parameters, α + β = 1 and α, β∈[0,1]; when the actual traffic volume of the road section is small, the interference between vehicles is small, the running speed is high, the traffic state is good, and the road section state value is also smaller, and vice versa;

[0124] The intersection traffic congestion state model is as follows:

[0125]

[0126] Among them, P i I is the traffic state value of intersection i; saturation R is the ratio of the actual traffic volume to the traffic capacity of the intersection, and its formula is: Where R is the intersection entrance saturation, Q and C are the actual traffic flow and traffic capacity of the entrance, S is the signal phase saturation flow, c is the signal cycle length, and g is the effective green light time; R i is the maximum saturation of the entrance road of the key section of the intersection; m i is the number of key sections at intersection i;

[0127] According to the regional road network topology, the node adjacency matrix A(G) that represents the physical connection relationship of the actual road network can be obtained. mn , the elements are expressed by the following formula:

[0128]

[0129] Among them, a ij =1 indicates that there are actual passable sections at intersections i and j, otherwise they do not exist or are prohibited;

[0130] By collecting traffic parameters on different road sections, the traffic flow of different road sections and intersections in [t k ,t k+1 ] time period; Based on the adjacency matrix and the traffic status of the road section, the road network adjacency traffic status matrix is constructed, which is recorded as A(P)={p ij (t)}, where p ij (t) can be expressed by the following formula:

[0131]

[0132] where p ij (t) represents the traffic status index of different road sections in the road network at time t; if p ij (t) = ∞ means the road section does not exist or is prohibited;

[0133] Step 22: Considering the traffic capacity per unit length, the spatial importance coefficient is used to reflect the spatial distribution of frequently congested road sections and intersections in the road network. The spatial importance coefficients of the road sections and intersections are as follows:

[0134]

[0135]

[0136] in, and is the key coefficient, and its value range is [0,1]; m iThe road section governed by the intersection;

[0137] In addition to paying attention to frequently congested sections and intersections in the spatial dimension, frequently congested conditions in the temporal dimension cannot be ignored either. Fully understanding the overall operating conditions of the urban road network in space and time is conducive to accurately planning traffic light control strategies, thereby improving road capacity and alleviating road congestion. Based on this, the percentage of congestion time is used to reflect the frequency and degree of congestion in the traffic network. The time importance coefficient is defined as the frequency coefficient of a certain degree of congestion in the urban road network within a certain time period of the day. Since the standards for judging traffic congestion in different time periods are not uniform, it is necessary to select different traffic congestion thresholds γ according to the actual situation of the city. If the congestion level f of the road section is t (l ij ) is greater than the traffic congestion threshold, it means it is in a congested state, otherwise it is in a smooth state. Each road section should meet the following conditions:

[0138]

[0139]

[0140] Among them, f t (l ij ) is used to determine the road section l in the tth time period within the statistical days ij The congestion state of Gra t (l ij ) represents the congestion level of a single road section in the tth time period within the statistical days, F ij (t) represents the road segment l ij The total number of frequencies with a congested state, N is the total number of time periods;

[0141] The time importance coefficient is expressed by the time period congestion duration ratio, and the road section time importance coefficient is expressed by the time period congestion duration ratio. and intersection time importance coefficient as follows:

[0142]

[0143]

[0144] Where F(t) is the sum of the frequencies of all road sections being congested;

[0145] The influence of the time importance coefficient and space importance coefficient of the above sections and intersections on the regional road network traffic status is normalized and sorted out to obtain the road network traffic status weight model; section weight model and intersection weight model The formulas are as follows:

[0146]

[0147]

[0148] Through the analysis of the spatiotemporal importance weight model of road sections and intersections in the regional road network, the road section weight coefficient matrix W L (t) and the intersection weight vector W I (t) are as follows:

[0149]

[0150] W I (t)=(w1 w2…w n )

[0151] Step 23: Based on the overall traffic state matrix of the regional road network obtained in step 21, the spatiotemporal importance weights of road sections and intersections are considered, and the regional road network state model is as follows:

[0152]

[0153] Among them, P(G) is the regional road network traffic state matrix; the number PI ij (t) is the traffic status of the road network unit. When i≠j, PI ij (t) represents the traffic status index of the road section. When i=j, PI ij (t) represents the state index of the intersection. When PI ij When (t)=∞, it means that there is no L ij Road section; PI ij The (t) value reflects the congestion level of a road section or intersection, with larger values indicating more congestion;

[0154] Calculate the value of the traffic congestion index PI of the regional road network as a whole:

[0155]

[0156] Among them, PI is the regional road network traffic congestion index, which is used to indicate the degree of traffic congestion in a sub-region of the road network. Analysis shows that the larger the PI value, the more congested the road network; the smaller the PI value, the smoother the traffic in the entire road network. λ is a model parameter with a value of [0,1].

[0157] The traffic congestion status model for each road section and intersection can reflect the congestion and distribution of traffic flow throughout the network, from which the main sources and directions of traffic flow can be inferred. The spatiotemporal importance coefficient weight model is derived through modeling and analysis based on both time and space. It can combine abundant traffic information resources to determine how traffic flow distribution changes over time and space across the entire network, providing a three-dimensional representation of the spatiotemporal traffic status of the regional road network. Finally, combining these two pieces of information, the overall traffic status matrix of the regional road network is calculated. This combines both macro-level road network traffic congestion status evaluation and micro-level traffic parameters, providing a comprehensive visualization method for observing and studying the evolution of regional road network traffic status.

[0158] Step three: Establish a traffic light control system based on a fuzzy neural network, use the regional road network traffic congestion index as the input of the fuzzy neural network, calculate the correspondence between the regional road network traffic congestion status and the traffic light control results over a period of time, obtain the predicted traffic light control information, and realize global adaptive signal control of multiple sections.

[0159] Fuzzy neural network adaptive control utilizes detectors installed at road sections or intersections to collect traffic data in real time and implements online optimal control based on this data. To achieve a more optimized signal timing strategy, this embodiment uses "three-dimensional traffic flow characteristic state information" and "temporal and spatial importance level information" as the control system inputs, and "traffic light control" (achieving the minimum value of this parameter is the control target of the signal timing system) as the control system output. The three-dimensional traffic flow characteristic state information is fuzzified to include four states: "very smooth, smooth, congested, and very congested," and the temporal and spatial importance level information includes four states: "very low, low, high, and very high," totaling 16 input states. Based on the adaptive signal control system, the corresponding signal control rules are implemented using a fuzzy neural network control algorithm, resulting in traffic light control for five states: "very short, short, medium, long, and very long."

[0160] The fuzzy neural network acquires the basic information for vehicle adaptive control rules by learning as much vehicle traffic status information as possible, thereby achieving global adaptive signal control for multiple road sections. Based on the general rules for setting fuzzy control rules and the basic traffic conditions that may occur in signal control, the fuzzy control rules for adaptive signal timing in this embodiment are set as shown in Table 1:

[0161] Table 1 Fuzzy control rules of self-organizing fuzzy neural network model

[0162]

[0163] The triangular membership function is used for traffic status analysis, and its function form is as follows:

[0164]

[0165] The specific form and distribution are determined by the parameters [a, b, c], requiring a≤b≤c. The parameter values are generally given based on expert experience scores.

[0166] The regional road network traffic congestion index PI changes according to the time period and is considered as a segment of sequence data. Therefore, this embodiment adds two modules, a sequence identifier and a sequence locator, to the fuzzy neural network structure to obtain a self-organizing fuzzy neural network model, wherein the sequence identifier module is used to receive the initial sample of the sequence to identify the sequence pattern in different learning sequences, and the sequence locator is used to locate the current state in the sequence. The two parts are used as inputs of the fuzzy rules to calculate the next sample of the identified sequence. Figure 4 As shown, the self-organizing fuzzy neural network model includes a discrimination layer, a sequence fuzzy layer, a storage layer, a sample fuzzy layer, a fuzzy rule layer, a normalization layer, and an output layer, and switches S1, S2, and S3 are set to manage the model. The specific functions of each structural layer are as follows:

[0167] Identification layer: used to calculate the identification number of each sequence. This layer accepts T sequence samples of the traffic state sequence. When x(t) is the input at the tth moment, the output of the pth neuron is expressed as:

[0168]

[0169] Among them, after T samples, the output of the p-th neuron is

[0170] Sequential fuzzy layer: It is used to receive the vector containing all the output information of the discriminator layer and define multivariate fuzzy sets for different identification numbers in the discriminator layer. The output of the pth neuron in the sequential fuzzy layer is expressed as:

[0171]

[0172] Among them, c1 p is the vector of the center of the pth fuzzy set, ∑1 is the diagonal matrix containing the width of the fuzzy set for each input dimension;

[0173] Storage layer: Each neuron in this layer is a discrete low-pass filter with a different time constant, λ p =p / (p+1), the output is a vector containing the current sample and d previous samples of the sequence, the formula is as follows:

[0174]

[0175] Sample fuzzy layer: The principle is similar to the sequence fuzzy layer. The neuron output formula is as follows:

[0176]

[0177] Fuzzy rule layer: Each neuron in this layer represents a fuzzy rule. The T-Norm operator is applied to calculate the membership based on the output of the sample fuzzy layer. The membership function of the pth fuzzy rule is defined as:

[0178]

[0179] Among them, μ p It is the membership degree of the current sequence to the pth fuzzy rule and the current position of its corresponding fuzzy rule.

[0180] Normalization layer: Normalize the calculated membership. The output of the p-th neuron in this layer is as follows:

[0181]

[0182] Among them, φ p is the normalized membership of the current sequence to the pth fuzzy rule and the current position of its corresponding fuzzy rule, and R is the number of fuzzy rules.

[0183] Output layer: The output of this layer is the inner product of the input vector and the weight vector W. The weight vector serves as the consequent parameter. The final output of the network is as follows:

[0184] o (7) =y(t)=x(t+1)=W T o (6)

[0185] The self-organizing fuzzy neural network model sequence constructed based on the above 7-layer structure has dynamic learning characteristics. During the learning process, switch S1 is closed and S2 is opened. For the first T sequence samples, switch S3 is closed. The model input is the original sequence sample x(t); after receiving T samples, switches S1 and S3 are opened and S2 is closed. The model uses the output of the storage layer as the first fuzzy set center of the fuzzy set layer, locates the current position of the sequence based on global information, and combines fuzzy rules to obtain the predicted sample.

[0186] The regional road network traffic status data is serialized based on time parameters and used as the input of a self-organizing fuzzy neural network model. The sequence rules of a specific time period are learned according to the time parameters, and the sequence rules are applied to a fuzzy neural network with adaptive function to obtain the traffic congestion status prediction value of the corresponding road section, intersection or regional section; and the predicted traffic light control information is obtained based on the traffic congestion status prediction value obtained above.

[0187] The above description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution, modification, or other variation to the technical solution and technical contents disclosed in the present invention shall be deemed to be within the scope of the present invention and still fall within the scope of protection of the present invention.

Claims

1. A traffic light control method based on multi-channel vehicle detection and stereo feature annotation, characterized in that: The specific steps include: Step a: Using a vehicle detection model to perform feature detection and collect vehicle information at the intersection; Step b: analyzing the vehicle congestion status on the three-dimensional plane road based on the intersection vehicle information collected in step a, adding the importance weight coefficients of each road section and intersection in the road network, obtaining the regional road network traffic status congestion index, and forming a three-dimensional feature annotation of the road network traffic flow; The process of adding the importance weight coefficients of each road section and intersection in the road network to obtain the regional road network traffic congestion index in step b specifically includes: The road network is represented by graph theory, which abstracts the road network into a weighted directed graph, abstracts the intersections into nodes in the road network, and abstracts the road segments between adjacent nodes into edges. The regional road network model composed of intersections and road segments can be described as: G = {I, L}; where I represents the set of intersections in the regional road network, I = {I1, I2, ... I n }, I n Indicates the nth intersection in the regional road network intersection set, L = {l ij |l ij =(I i ,I j ); i, j = 1, 2, ..., n, i ≠ j} represents the set of road segments between adjacent nodes; l ij Indicates that there is a road from intersection I i To intersection I j The road section has a length of l ij ; Step b1: First, establish the traffic state model of the road section and intersection, and then calculate and integrate it into the traffic state matrix of the entire regional road network; The road section traffic state model is as follows: in, For section L ij Traffic status value, For section L ij The maximum speed limit, v ij For section L ij The average speed, q ij is the measured flow rate, C ij is the traffic capacity, α and β are model parameters, α + β = 1 and α, β∈[0,1]; when the actual traffic volume of the road section is small, the interference between vehicles is small, the running speed is high, the traffic state is good, and the road section state value is also smaller, and vice versa; The intersection traffic state model is as follows: Among them, P i I is the traffic state value of intersection i; saturation R is the ratio of the actual traffic volume to the traffic capacity of the intersection, and its formula is: Where R is the intersection entrance saturation, Q and C are the actual traffic flow and traffic capacity of the entrance, S is the signal phase saturation flow, c is the signal cycle length, and g is the effective green light time; R i is the maximum saturation of the entrance road of the key section of the intersection; m i is the number of key sections at intersection i; According to the regional road network topology, the node adjacency matrix A(G) that represents the physical connection relationship of the actual road network can be obtained. mn , the elements are expressed by the following formula: Among them, a ij =1 indicates that there are actual passable sections at intersections i and j, otherwise they do not exist or are prohibited; By collecting traffic parameters on different road sections, the traffic flow of different road sections and intersections in [t k ,t k+1 ] time period; Based on the adjacency matrix and the traffic status of the road section, the road network adjacency traffic status matrix is constructed, which is recorded as A(P)={p ij (t)}, where p ij (t) can be expressed by the following formula: where p ij (t) represents the traffic status index of different road sections in the road network at time t; if p ij (t) = ∞ means the road section does not exist or is prohibited; Step b2: Considering the traffic capacity per unit length, the spatial importance coefficient is used to reflect the spatial distribution of frequently congested sections and intersections in the road network. The spatial importance coefficients of the sections and intersections are as follows: in, and is the key coefficient, and its value range is [0,1]; m i The road section governed by the intersection; The congestion time percentage is used to reflect the frequency and degree of traffic network congestion. The time importance coefficient is defined as the frequency coefficient of a certain degree of congestion in a certain time period of a day in the urban road network. The standards for judging congestion in different time periods are not uniform, so different traffic congestion thresholds γ need to be selected according to the actual situation of the city. If the congestion level f of a road section is t (l ij ) is greater than the traffic congestion threshold, it indicates a congested state, otherwise it is in a smooth state; each road section should meet the following conditions: Among them, f t (l ij ) is used to determine the road section l in the tth time period within the statistical days ij The congestion state of Gra t (l ij ) represents the congestion level of a single road section in the tth time period within the statistical days, F ij (t) represents the road segment l ij The total number of frequencies with a congested state, N is the total number of time periods; The time importance coefficient is expressed by the time period congestion duration ratio, and the road section time importance coefficient is expressed by the time period congestion duration ratio. and intersection time importance coefficient as follows: Where F(t) is the sum of the frequencies of all road sections being congested; The influence of the time importance coefficient and space importance coefficient of the above sections and intersections on the regional road network traffic status is normalized and sorted out to obtain the road network traffic status weight model; section weight model and intersection weight model The formula is as follows: Through the analysis of the spatiotemporal importance weight model of road sections and intersections in the regional road network, the road section weight coefficient matrix W L (t) and the intersection weight vector W I (t) are as follows: W I (t)=(w1 w2 … w n ); Step b3: Based on the traffic state matrix of the regional road network obtained in step b1, the spatiotemporal importance weights of road sections and intersections are considered, and the regional road network state model is as follows: Among them, P(G) is the regional road network traffic state matrix; the number PI ij (t) is the traffic status of the road network unit. When i≠j, PI ij (t) represents the traffic status index of the road section. When i=j, PI ij (t) represents the state index of the intersection. When PI ij When (t)=∞, it means that there is no L ij Road section; PI ij The (t) value reflects the congestion level of a road section or intersection, with larger values indicating more congestion; Calculate the value of the traffic congestion index PI of the regional road network as a whole: Among them, PI is the regional road network traffic congestion index, which is used to indicate the degree of traffic congestion in a sub-region of the road network. Through analysis, it can be seen that the larger the PI value, the more congested the road network; the smaller the PI value, the smoother the traffic in the entire road network. λ is a model parameter with a value of [0,1]. Step c: Establish a traffic light control system based on a fuzzy neural network. Use the regional road network traffic congestion index obtained in step b as the input of the fuzzy neural network to calculate the corresponding relationship between the regional road network traffic congestion status and the traffic light control results over a period of time, and obtain the predicted traffic light control information.

2. The traffic light control method based on multi-channel vehicle detection and three-dimensional feature annotation according to claim 1 is characterized in that: The vehicle detection model is a vehicle detection model that integrates a multi-channel high-resolution network with an attention mechanism. The model is divided into a road network perception layer, a training network layer, and a feature output layer. The road network perception layer uses a combination of cameras and radar sensors and operates 24 / 7 in zero light and various weather conditions to implement vehicle detection, lane line recognition, vehicle distribution, and time recording feature detection. The collected road information is passed to the feature output layer, and digital information of each monitored intersection is extracted as input to the training network layer. The feature fusion method of integrating the attention mechanism in the high-resolution network is to use the feature connection weight and weight threshold at the network feature fusion connection to perform multi-resolution feature fusion on multiple adjacent network paths in the parallel network structure to obtain a high-resolution parallel network module, where the weight threshold is a hyperparameter and the feature connection weight is a learnable weight parameter; during the network model training process, the feature connection weight is adaptively learned based on the loss of reverse transmission to express the degree of influence of the current feature map on the network.

3. The traffic light control method based on multi-channel vehicle detection and three-dimensional feature annotation according to claim 2 is characterized in that: A feature fusion method based on skip connections is introduced for high-resolution networks, and skip connections are used to perform cross-layer feature fusion on feature maps at symmetrical positions in each network path in the parallel network structure.

4. The traffic light control method based on multi-channel vehicle detection and three-dimensional feature labeling according to claim 2 is characterized in that: The specific steps of multi-resolution feature fusion of the feature map in the high-resolution parallel network module are as follows: let the feature map of the q+1th convolution layer in the p+1 network path in the high-resolution parallel network module be Feature (p+1)(q+1) , where p = 1, 2, ..., R and q = 1, 2, ..., Q p , R is the total number of network paths in the high-resolution cascade network module, Q p is the total number of convolutional layers in the p-th network path, then Feature (p+1)(q+1) The steps of multi-resolution feature fusion are as follows: Step a1: Get feature map pq 、Feature p(q+1) and Feature p(q+2) , and obtain the feature connection weight W corresponding to the feature map pq 、W p(q+1) and W p(q+2) ; Among them, Feature pq Represents the feature map of the qth convolutional layer in the pth network path in the high-resolution parallel network module, Feature p(q+1) Indicates the pth network path q+1 The feature map of the convolution layer, Feature p(q+2) Indicates the pth network path q+2 The feature map of the convolutional layer, W pq Represents Feature pq The corresponding feature connection weight, W p(q+1) Represents Feature p(q+1) The corresponding feature connection weight, W p(q+2) Represents Feature p(q+2) The corresponding feature connection weights; Step a2: Connect the features with weights W in sequence pq 、W p(q+1) 、W p(q+2) Compared with the weight threshold K, the connection activity parameter is obtained; the feature connection weight W pq Connection activity parameter For example: Multi-resolution feature fusion is performed based on the connection activity parameters, and the following is obtained: When the connection weight of a feature is lower than the weight threshold, the feature map corresponding to the feature connection weight will be inactivated, and there is no need to fusion the inactivated feature map in the subsequent multi-resolution feature fusion.

5. The traffic light control method based on multi-channel vehicle detection and three-dimensional feature labeling according to claim 3 is characterized in that: The feature extraction part of the vehicle detection model uses cross-stage residual blocks to replace the traditional residual blocks in the feature extraction network. The feature fusion part introduces spatial pyramid pooling to fuse local area features with global area features, fully integrating and utilizing the feature information extracted from different deep network layers. After image preprocessing and lane line detection on the input image, a median-based background extraction method is used to subtract each frame from the calculated background image, and the subtracted image is binarized. A background difference algorithm is then used on the extracted background image to extract the foreground vehicle target. After the foreground vehicle target is extracted, the image is further segmented into a single lane image, and then feature extraction is performed on the single lane image using projection features. After extracting the single lane features, the K-nearest neighbor algorithm is used to classify the single lane feature vector to obtain the final detection result; The expression for the binarization processing of the image is as follows: Where x, y are the coordinate points of the corresponding image, B(x, y) is the background frame obtained by background extraction, and I i (x, y) represents the i-th frame image in the image sequence, D i (x,y) represents the result of the difference of point (x,y) on the i-th frame; if the difference between the two images at this point is greater than the threshold T, the result is 1, and if the difference is less than the threshold T, the result is 0.

6. The traffic light control method based on multi-channel vehicle detection and three-dimensional feature annotation according to claim 1 is characterized in that: A self-organizing fuzzy neural network model is obtained by adding two modules, a sequence identifier and a sequence locator, to the fuzzy neural network structure, wherein the sequence identifier module is used to receive the initial sample of the sequence to identify the sequence pattern in different learning sequences, and the sequence locator module is used to locate the current state in the sequence, and the two parts are used as inputs of the fuzzy rules to calculate the next sample of the identified sequence.

7. The traffic light control method based on multi-channel vehicle detection and three-dimensional feature annotation according to claim 6 is characterized in that: The self-organizing fuzzy neural network model includes a discrimination layer, a sequence fuzzy layer, a storage layer, a sample fuzzy layer, a fuzzy rule layer, a normalization layer, and an output layer. Switches S1, S2, and S3 are set to manage the model. The specific functions of each structural layer are as follows: (1) Identification layer: used to calculate the identification number of each sequence. This layer accepts T sequence samples of the traffic state sequence. When x(t) is the input at the tth moment, the output of the pth neuron is expressed as: Among them, after T samples, the output of the p-th neuron is (2) Sequential fuzzy layer: It is used to receive the vector containing all the output information of the discrimination layer and define multivariate fuzzy sets for different identification numbers in the discrimination layer. The output of the pth neuron in the sequential fuzzy layer is expressed as: Among them, c1 p is the vector of the center of the pth fuzzy set, ∑1 is the diagonal matrix containing the width of the fuzzy set for each input dimension; (3) Storage layer: Each neuron in this layer is a discrete low-pass filter with different time constants, λ p =p / (p+1), the output is a vector containing the current sample and d previous samples of the sequence, the formula is as follows: (4) Sample fuzzy layer: The principle is similar to the sequence fuzzy layer. The neuron output formula is as follows: (5) Fuzzy rule layer: Each neuron in this layer represents a fuzzy rule. The T-Norm operator is applied to calculate the membership based on the output of the sample fuzzy layer. The membership function of the pth fuzzy rule is defined as: Among them, μ p is the membership degree of the current sequence to the pth fuzzy rule and the current position of its corresponding fuzzy rule; (6) Normalization layer: Normalize the calculated membership. The output of the p-th neuron in this layer is as follows: Among them, φ p is the normalized membership of the current sequence to the pth fuzzy rule and the current position of its corresponding fuzzy rule, and R is the number of fuzzy rules; (7) Output layer: The output of this layer is the inner product of the input vector and the weight vector W. The weight vector serves as the consequent parameter. The final output of the network is as follows: o (7) =y(t)=x(t+1)=W T ro (6) ; During the learning process, switch S1 is closed and S2 is open. For the first T sequence samples, switch S3 is closed, and the model input is the original sequence sample x(t). After receiving T samples, switches S1 and S3 are opened, and S2 is closed. The model uses the output of the storage layer as the first fuzzy set center of the fuzzy set layer, locates the current position of the sequence based on global information, and combines fuzzy rules to obtain the predicted sample. The regional road network traffic status congestion index obtained in step b is serialized based on the time parameter and used as the input of the fuzzy neural network model. The sequence law of the specific time period is learned according to the time parameter, and the sequence law is applied to the fuzzy neural network with adaptive function to obtain the traffic congestion status prediction value of the corresponding road section, intersection or regional road section. Based on this, the predicted traffic light control information is obtained for the traffic congestion status prediction value obtained above.

8. The traffic light control method based on multi-channel vehicle detection and three-dimensional feature labeling according to claim 7 is characterized in that: The fuzzy control rules of the self-organizing fuzzy neural network model are as follows: "traffic flow three-dimensional characteristic state information" and "temporal and spatial importance level information" are used as inputs to the traffic light control system, and "traffic light control" is used as the output of the traffic light control system. The fuzzy three-dimensional characteristic state information of traffic flow includes "very smooth, smooth, crowded, very crowded", and the temporal and spatial importance level information includes "very low, low, high, very high". The two together constitute a total of 16 input states, and ultimately the traffic light control is achieved in five states: "very short, short, medium, long, very long".

9. The traffic light control method based on multi-channel vehicle detection and three-dimensional feature annotation according to claim 7 is characterized in that: The fuzzy membership function adopts a triangular membership function, and its function form is as follows: The specific form and distribution are determined by the parameters [a, b, c], requiring a≤b≤c, and the parameter values are given based on expert experience scores.

Images