Traffic signal control optimization method and system based on big data and deep learning

Abstract

The present invention discloses a traffic signal control optimization method and system based on big data and deep learning, belonging to the field of traffic control technology. The method includes: obtaining historical traffic basic information and historical traffic flow information within the control area, dividing the historical traffic basic information and historical traffic flow information into training sets and test sets, and preprocessing the historical traffic flow information to obtain processed data; dividing the traffic control sub-areas based on the historical traffic basic information; constructing a traffic control model, and performing signal control on the traffic control sub-areas based on the processed data of the training set and the traffic control model; constructing a virtual traffic operation model, and performing a virtual operation of traffic signal control based on the processed data of the test set and the traffic control model to obtain traffic signal control results; and quantitatively evaluating the traffic signal control results to complete traffic signal control. The present invention provides an efficient and intelligent solution for urban traffic management.

Description

Technical Field

[0001] The present invention belongs to the field of traffic control technology, and specifically relates to a control optimization method and system for traffic signals based on big data and deep learning. Background Art

[0002] Traffic signal control is a crucial area of ​​urban traffic management, directly impacting road efficiency, congestion relief, and energy consumption, and playing a key role in improving the quality of urban life. Traditional traffic signal control methods, which rely on fixed-cycle timing or simple sensor-based adjustments, struggle to cope with complex and changing traffic patterns. These methods often exhibit slow response and rigid timing during peak hours, emergencies, or changes in regional traffic demand, leading to inefficient use of road resources and even exacerbating congestion.

[0003] Against this backdrop, the intelligent upgrade of traffic signal control systems has become a pressing issue. The primary challenge lies in extracting valuable information from massive amounts of traffic data. Urban traffic data encompasses multiple dimensions, including vehicle volume, travel time, and congestion levels. The data is massive and dynamically changing. Without the ability to efficiently process and analyze these data, signal control systems will struggle to capture real-time trends in traffic patterns. A deeper problem stemming from this is the difficulty in adaptively adjusting signal timing plans. Traditional systems often employ preset rules and lack the ability to dynamically learn about traffic conditions, resulting in the inability to flexibly optimize signal cycles based on actual needs, which in turn limits the evolution of control strategies.

[0004] Therefore, how to achieve dynamic optimization and adaptive adjustment of traffic signal timing through big data analysis has become a key issue in the intelligent upgrade of traffic signal control. Summary of the Invention

[0005] The present invention aims to address the deficiencies of the prior art and proposes a traffic signal control optimization method based on big data and deep learning. The method comprises the following steps:

[0006] S1. Obtain historical basic traffic information and historical traffic flow information within the control area, divide the historical basic traffic information and the historical traffic flow information into a training set and a test set, and preprocess the historical traffic flow information to obtain processed data;

[0007] S2. Dividing traffic control sub-areas based on the historical traffic basic information;

[0008] S3. Build a traffic control model, and perform signal control on the traffic control sub-area based on the processed data of the training set and the traffic control model;

[0009] S4. Construct a virtual traffic operation model, perform a virtual traffic signal control operation based on the processed data of the test set and the traffic control model, and obtain a traffic signal control result;

[0010] S5. Quantitatively evaluate the traffic signal control result to complete the traffic signal control.

[0011] Further preferably, in S1, the historical traffic flow information includes video information collected by traffic cameras; the processed data includes: vehicle information and pedestrian information;

[0012] The method for obtaining the processed data includes:

[0013] S12, acquiring the video information, collecting frame data of the video information, and obtaining a traffic flow image;

[0014] S13, dividing the traffic flow image into a lane area and a pedestrian waiting area;

[0015] S14, performing vehicle identification on the lane area to obtain vehicle information;

[0016] S15. Perform pedestrian recognition in the pedestrian waiting area to obtain pedestrian information.

[0017] Further preferably, in S13, the method of dividing the lane area and the pedestrian waiting area includes:

[0018] Detecting lane lines in the traffic flow image using a lane line detection model, and dividing the lane area based on the detection results;

[0019] The positions of pedestrians are identified based on the traffic flow image, and the pedestrian waiting area is divided based on the positions of pedestrians.

[0020] Further preferably, the lane line detection model includes: an encoder part and a decoder part;

[0021] The encoder part includes: a first convolution module and a second convolution module; the first convolution module includes three first convolution layers; the second convolution module includes two second convolution layers;

[0022] The first convolutional layer includes: 3×3 convolution, batch normalization and ReLU activation; a residual block is set between the two first convolutional layers, and the residual block adopts 1×1 convolution;

[0023] The second convolution module includes two branches: a first branch and a second branch, each of the first branch and the second branch includes two nonlinear activation functions, and the nonlinear activation function of the second branch is preceded and followed by two basis functions with a kernel size of 1; the outputs of the two branches are then added;

[0024] The decoder part includes: the second convolution module and the third convolution module;

[0025] The third convolution module includes three third convolution layers; the third convolution layers include: 3×3 convolution, batch normalization and ReLU activation;

[0026] An attention connection module is provided between the decoder part and the encoder part.

[0027] Further preferably, in S2, the method of dividing the traffic control sub-areas includes:

[0028] S21, extracting road sections and intersections based on the historical traffic basic information;

[0029] S22, calculating the correlation between the road sections to obtain a correlation matrix;

[0030] S23, clustering the road sections based on the association matrix to obtain a first cluster set;

[0031] S24. Calculate the correlation coefficient between the first clusters, merge the first clusters based on the correlation coefficient to obtain a second cluster; and complete the division of the traffic control sub-area based on the second cluster.

[0032] The calculation method of the correlation coefficient includes:

[0033] ,

[0034] ,

[0035] X = RI × βRC ;

[0036] Where, RI represents the first correlation coefficient, RI ( m,n ) represents a cluster m and clusters n The first correlation coefficient between RC represents the second correlation coefficient, RC ( m 1 ,n 1) Represents a subcluster m 1 and subclusters n The second correlation coefficient between 1; X represents the correlation coefficient; EC ( m , n ) represents a connected cluster m and clusters nThe sum of the weights of all edges; EC ( m )and EC ( n ) represent clusters m and clusters n The edge weights that are divided into two parts of approximately equal size; Indicates that the cluster C Divide into subclusters m 1 and subclusters n An edge weight of 1; and Respectively represent the sub-clusters m 1 and subclusters n 1 is the edge weight of the partition into two parts of approximately equal size; β Represents the weight of the second correlation coefficient.

[0037] Further preferably, the method for obtaining the correlation matrix includes:

[0038] Constructing a road network model according to the road sections and the intersections;

[0039] Calculating correlation coefficients between the intersections, and obtaining the correlation matrix based on the correlation coefficients and the road network model;

[0040] The calculation method of the correlation coefficient includes:

[0041] ,

[0042] Where, FQ ( A , B ) represents the correlation coefficient; n 0 indicates the number of traffic branches in the two intersection directions; QAmax Indicates the maximum flow value of the traffic branch; N Indicates the number of lanes; Qk ( A → B ) indicates the k Traffic flow value of the branch road, tAB Indicates that the vehicle is passing through the intersection A , B The average travel time between.

[0043] Further preferably, the traffic control model includes: a flow prediction module, a decision module and an optimization module;

[0044] The traffic prediction module is used to predict traffic flow in the next time period based on historical traffic flow and real-time traffic flow;

[0045] The decision module performs traffic control based on the predicted traffic flow to obtain a traffic control strategy;

[0046] The optimization module is used to optimize the result of the traffic control strategy and perform signal control on the traffic control sub-area based on the optimized traffic control strategy.

[0047] Further preferably, the method for constructing the virtual traffic operation model includes:

[0048] Constructing a dynamic database and a static database; the dynamic database is used to store pedestrian and vehicle models; the static database is used to store road and traffic facility models;

[0049] Setting a traffic operation scenario based on the historical traffic basic information of the test set and the processed data;

[0050] The traffic signal control result is obtained based on the optimized traffic control strategy output by the traffic operation scenario and the traffic control model.

[0051] Further preferably, a signal control quantitative evaluation index is used to quantitatively evaluate the traffic signal control result, and the signal control quantitative evaluation index includes:

[0052] ,

[0053] Where, Score Represents the credit control quantitative evaluation index; BS An equilibrium index indicating the degree of saturation between phases; v Indicates the number of intersection phases; Indicates the signal control quantitative evaluation index of the phase; α represents the weight of pedestrian tolerance in the traffic control quantitative evaluation index; η Indicates pedestrian tolerance;

[0054] in,

[0055] ,

[0056] ,

[0057] Where, ROR 、 、 They represent red light occupancy, queue ratio and green light utilization respectively; 、 、 are the weight coefficients of red light occupancy rate, queue ratio and green light utilization rate respectively; Indicates phase i saturation; Indicates the number of intersection phases.

[0058] The present invention also provides a traffic signal control optimization system based on big data and deep learning, including: an information collection module, an area division module, a signal control module, a virtual operation module, and an evaluation module;

[0059] The information collection module is used to obtain historical traffic basic information and historical traffic flow information in the control area, divide the historical traffic basic information and the historical traffic flow information into training sets and test sets, and pre-process the historical traffic flow information to obtain processed data;

[0060] The area division module is used to divide the traffic control sub-areas based on the historical traffic basic information;

[0061] The signal control module is used to construct a traffic control model and perform signal control on the traffic control sub-area based on the processed data of the training set and the traffic control model;

[0062] The virtual operation module is used to build a virtual traffic operation model, perform a virtual operation of traffic signal control based on the processed data of the test set and the traffic control model, and obtain a traffic signal control result;

[0063] The evaluation module is used to perform quantitative evaluation on the traffic signal control result to complete the control of the traffic signal.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1. Accurate data collection and dynamic processing capabilities:

[0066] By collecting real-time video information from traffic cameras and combining it with an improved lane detection model (encoder-decoder architecture and attention connection module) and pedestrian recognition technology, this method can efficiently delineate lane areas from pedestrian waiting areas, accurately extracting multi-dimensional data such as vehicle position, speed, and pedestrian count. Compared to traditional fixed sensors or manual counting, this method significantly improves data coverage and recognition accuracy, providing a reliable basis for subsequent dynamic control.

[0067] 2. Adaptive traffic sub-area division:

[0068] Based on the road network topology correlation matrix and a cluster merging algorithm, traffic control sub-areas are dynamically divided. By calculating the association and correlation coefficient of road sections, highly correlated clusters are scientifically merged, avoiding the response lag caused by traditional fixed zoning. This method can flexibly respond to complex scenarios such as peak hours and emergencies, achieve inter-regional collaborative optimization, and improve overall road network resource utilization.

[0069] 3. Intelligent prediction and dynamic optimization control:

[0070] The traffic control model integrates flow prediction, decision-making, and optimization modules. It uses historical data and real-time traffic flow to predict future traffic demand and dynamically adjusts signal cycle and phase timing using a nonlinear optimization function. Compared to traditional fixed-cycle or simple sensor-based control, this approach significantly reduces average intersection delays, effectively alleviating congestion and improving traffic efficiency.

[0071] 4. Real-time quantitative evaluation and closed-loop feedback:

[0072] The system uses a quantitative signal control evaluation index to assess control effectiveness in real time, combined with a threshold-triggered strategy adjustment mechanism to form a closed-loop feedback loop of "acquisition-decision-optimization-evaluation." This system continuously optimizes signal control solutions, avoiding the strategy rigidity that can occur in traditional systems due to their lack of dynamic learning capabilities.

[0073] 5. Resource conservation and environmental friendliness:

[0074] By dynamically optimizing vehicle start and stop times and waiting times, this system can reduce fuel consumption and exhaust emissions, while improving pedestrian crossing safety and providing technical support for the sustainable development of smart cities.

[0075] In summary, the present invention solves the core problems of traditional signal control systems such as slow response, rigid timing, and poor regional coordination through the deep integration of big data-driven, deep learning, and dynamic optimization technologies, providing an efficient and intelligent solution for urban traffic management. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0077] Figure 1 This is a flow chart of a traffic signal control optimization method based on big data and deep learning according to an embodiment of the present invention;

[0078] Figure 2 This is a schematic diagram of the preprocessing process according to an embodiment of the present invention;

[0079] Figure 3 The figure is a schematic diagram of the process of dividing traffic control sub-areas according to an embodiment of the present invention. DETAILED DESCRIPTION

[0080] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0081] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0082] Example 1:

[0083] like Figure 1 As shown, this embodiment provides a traffic signal control optimization method based on big data and deep learning, and the method includes the following steps:

[0084] S1. Obtain historical traffic basic information and historical traffic flow information within the control area, divide the historical traffic basic information and historical traffic flow information into training sets and test sets, and pre-process the historical traffic flow information to obtain processed data. In this embodiment, the historical traffic flow information includes video information collected by traffic cameras; the processed data includes: vehicle information and pedestrian information. Figure 2 As shown, the method for obtaining the processed data includes:

[0085] S12: Obtain video information, collect frame data of the video information, and obtain a traffic flow image.

[0086] S13: Divide the traffic flow image into a lane area and a pedestrian waiting area.

[0087] Specifically, the method of dividing the lane area and the pedestrian waiting area includes:

[0088] The lane line detection model is used to detect the lane lines in the traffic flow image, and the lane area is divided based on the detection results.

[0089] The lane detection model further comprises an encoder and a decoder. The encoder comprises a first convolutional module and a second convolutional module. The first convolutional module comprises three first convolutional layers. The second convolutional module comprises two second convolutional layers. The first convolutional layer comprises 3×3 convolution, batch normalization, and ReLU activation. A residual block is provided between the two first convolutional layers, and the residual block uses 1×1 convolution. The input or downsampled features are added to the current feature map to achieve an efficient residual connection. The second convolutional module comprises two branches: a first branch and a second branch. The first and second branches each comprise two nonlinear activation functions. The nonlinear activation function of the second branch is preceded and followed by two basis functions with a kernel size of 1. The outputs of the two branches are then added. The decoder comprises a second convolutional module and a third convolutional module. The third convolutional module comprises three third convolutional layers. The third convolutional layer comprises 3×3 convolution, batch normalization, and ReLU activation. An attention connection module is provided between the decoder and encoder. Among them, the attention connection module is improved on the basis of the traditional ELA module. The traditional ELA module uses avg pool in the spatial dimension to obtain the feature vectors in the horizontal and vertical directions. The attention connection module of this embodiment combines avg pool and max pool in the spatial dimension, and max pool is used to obtain the most significant feature values ​​in the local area. The avg pool and max pool are used to obtain the feature vectors in the horizontal and vertical directions respectively, and the obtained feature vectors are spliced. The lane line detection model is optimized by collecting scene images in different situations to obtain the optimal detection model and realize the detection of lane lines. The lane area is divided by the detected lane lines.

[0090] Pedestrian positions are identified based on traffic flow images, and pedestrian waiting areas are divided based on pedestrian positions.

[0091] In this embodiment, the non-lane area is divided into a pedestrian waiting area.

[0092] S14: Identify vehicles in the lane area to obtain vehicle information.

[0093] In this embodiment, a vehicle recognition model is used to identify vehicles in a lane area image, and the vehicle's position and average speed over the 10 consecutive seconds before the vehicle stops are determined. The vehicle recognition model includes a vehicle recognition module and a calculation module; the recognition module is used to identify vehicles in the lane area image, and the calculation module is used to calculate speed information for the identified vehicles.

[0094] Specifically, the vehicle recognition module includes a backbone network, a neck network, and a head detection network. The backbone network includes a first backbone module, a second backbone module, a third backbone module, and a fourth backbone module; the neck network includes a first neck module, a second neck module, and a third neck module; and the head detection network includes a detection module. In this embodiment, the first backbone module consists of a convolutional layer (Conv); the input lane area image passes through a 1*1 convolutional layer to extract convolutional features. The second backbone module includes four groups of second backbone units, each group of second backbone units consists of a guide block and an aggregation block; the guide block consists of four feature extraction blocks, namely the first feature extraction block, the second feature extraction block, the third feature extraction block, and the fourth feature extraction block; the obtained convolutional features are respectively input into the first feature extraction block and the second feature extraction block to learn local features and contextual features, thereby obtaining local features and contextual features. In this embodiment, the first feature extraction block consists of a 3*3 standard convolutional layer, and the second feature extraction block consists of a 3*3 depthwise convolution. The resulting local and contextual features are then fed into the third feature extraction module for fusion, yielding the third fused feature. The third feature extraction module consists of a concatenation layer, a batch normalization (BN) layer, and a ReLU activation function. The concatenation layer fuses the local and contextual features, and batch normalization and activation further enhance the feature representation, yielding the third fused feature. The third fused feature is then fed into the fourth feature extraction block, yielding the fourth fused feature. The fourth feature extraction block consists of a global average pooling layer and two fully connected layers. The global average pooling layer aggregates contextual information from the third fused feature, and the feature is then further processed by two fully connected layers to yield the fourth fused feature. The aggregation block consists of two branches: the first and second branches. The fourth fused feature is first channel-compressed by a 1x1 standard convolution and then fed into the first and second branches, respectively. The first branch consists of a reparameterized convolution, a 3x3 standard convolution, and a 1x1 standard convolution. The second branch processes the input using a 1x1 standard convolution. The outputs of the first and second branches are then fed into the concatenation layer for integration, followed by a 1x1 standard convolution. The third backbone module uses a standard spatial pyramid pooling module. The fourth backbone module uses a standard C2PSA module. The neck network includes: a first neck module, a second neck module, and a third neck module. The first neck module consists of a single 1x1 convolutional layer; the second neck module uses the same structure as the second backbone module; and the third neck module uses the same structure as above.The detection module adopts three detection branches, which are composed of P3, P4, and P5 detection layers. The P3, P4, and P5 detection layers are respectively connected to a 1*1 group normalization convolution layer. The group normalization convolution layer groups the feature dimensions of the input image and performs normalization within each group. It then passes through two DE convolution layers to capture more image details. The feature map output by the DE convolution layer then passes through the Conv_Reg and Conv_Cls layers for bounding box regression and category classification, and scale is used to adjust the scale of the final output. In this embodiment, the DE convolution layer includes five convolution layers, namely: ordinary convolution, center difference convolution, angular difference convolution, horizontal difference convolution, and vertical difference convolution. First, the center difference convolution and angular difference convolution are used to encode the prior information into the convolution layer, and then the horizontal difference convolution and vertical difference convolution are used to integrate the traditional local descriptor into the convolution layer, and the representation and generalization capabilities are enhanced by learning gradient information. Normal convolution and differential convolution are used to obtain horizontal intensity and enhance gradient information, and integrate the features learned by the five convolutional layers. Finally, reparameterization technology is used to obtain rich features without introducing additional computational cost, which is used as the output of the DE convolutional layer.

[0095] The calculation module tracks the identified vehicle and obtains the vehicle's speed information by combining the sampling time between the upper and lower frames, the vehicle's movement distance, and the image acquisition ratio. It calculates the average speed of the vehicle over a continuous 10-second period.

[0096] S15. Perform pedestrian recognition in the pedestrian waiting area to obtain pedestrian information.

[0097] In this embodiment, the pedestrian information includes: the number of pedestrians and the intention of the pedestrians. In this embodiment, the number of pedestrians and the intention of the pedestrians are identified by a pedestrian recognition model.

[0098] Specifically, the pedestrian recognition model includes: a pedestrian recognition unit and an intention recognition unit; wherein the pedestrian recognition unit adopts the structure of the vehicle recognition unit described above to identify pedestrians. The intention recognition unit is used to track the identified pedestrian and obtain the pedestrian's intention by combining the sampling time between the upper and lower frames and the pedestrian's movement distance. In this embodiment, the pedestrian's intention includes waiting and moving; by setting a motion threshold and comparing it with the obtained pedestrian's movement distance, when the pedestrian's movement distance is greater than the set operating threshold, the pedestrian's intention is judged to be moving, otherwise it is waiting. Pedestrian tolerance is also calculated based on pedestrians with moving intentions and pedestrians with waiting intentions.

[0099] S2. Divide traffic control sub-areas based on historical traffic basic information. Figure 3 As shown, the methods for dividing traffic control sub-areas include:

[0100] S21. Extract road sections and intersections based on basic traffic information.

[0101] S22. Calculate the correlation between road sections to obtain a correlation matrix.

[0102] A further implementation is that the method for obtaining the correlation matrix includes:

[0103] A road network model is constructed based on road sections and intersections; the correlation coefficients between intersections are calculated, and a correlation matrix is ​​obtained based on the correlation coefficients and the road network model.

[0104] Incorporating graph theory, the road network in the traffic control area is described as a topological structure graph. The extracted intersections are defined as points, and the road segments connecting them are defined as edges. Each segment is numbered and represented using an undirected graph as an adjacency matrix. In the adjacency matrix, the diagonal elements are set to 0, the matrix elements corresponding to the numbers of adjacent segments are set to 1, and the matrix elements corresponding to non-adjacent segments are set to 0, thus forming a road network model.

[0105] In this embodiment, the calculation method of the correlation coefficient includes:

[0106] ,

[0107] Where, F Q(A,B) represents the correlation coefficient; n 0 indicates the number of traffic branches in the two intersection directions; Q Amax Indicates the maximum flow value of the traffic branch; N Indicates the number of lanes; Q k（A→B） Indicates the k Traffic flow value of the branch road, t AB Indicates that the vehicle is passing through the intersection A , B The average travel time between.

[0108] S23. Cluster the road sections based on the association matrix to obtain a first cluster set.

[0109] First, randomly divide a traffic control subarea into a number of subareas. Arrange the local density of the first b road segments in descending order, and use the first b road segments as cluster centers. When cluster centers are adjacent in space, the cluster center with the highest local density is selected as the new cluster center. This process is repeated for all cluster centers until all cluster centers are non-adjacent, resulting in a new cluster point set. The shortest path between the remaining road segments and the new cluster point is calculated. Using this shortest path value, the remaining road segments are merged into the new cluster point until clustering is complete, resulting in a number of clusters.

[0110] S24 , calculating the correlation coefficients between the first clusters, merging the first clusters based on the correlation coefficients to obtain second clusters; and completing the division of the traffic control sub-areas based on the second clusters.

[0111] In this embodiment, the correlation coefficient is calculated by the following formula:

[0112] ,

[0113] ,

[0114] X = RI × βRC ;

[0115] Where, RI represents the first correlation coefficient, RI ( m,n ) represents a cluster m and clusters n The first correlation coefficient between RC represents the second correlation coefficient, RC ( m 1 ,n 1) Represents a subcluster m 1 and subclusters n The second correlation coefficient between 1; X represents the correlation coefficient; EC ( m , n ) represents a connected cluster m and clusters n The sum of the weights of all edges; EC ( m )and EC ( n ) represent clusters m and clusters n The edge weights that are divided into two parts of approximately equal size; Indicates that the cluster C Divide into subclusters m 1 and subclusters n An edge weight of 1; and Respectively represent the sub-clusters m 1 and subclusters n 1 is the edge weight of the partition into two parts of approximately equal size; β Represents the weight of the second correlation coefficient.

[0116] A merging threshold is set, and two clusters with correlation coefficients greater than the merging threshold are merged until all clusters are merged to obtain a second cluster set, which is then divided into corresponding traffic control sub-areas according to the number of clusters in the second cluster set.

[0117] S3. Build a traffic control model and perform signal control on the traffic control sub-area based on the processed data of the training set and the traffic control model.

[0118] Further implementation is that the traffic control model includes: a flow prediction module, a decision module and an optimization module; the flow prediction module is used to predict the flow based on historical flow and real-time flow and the next time period; the decision module performs traffic control based on the predicted flow and obtains a traffic control strategy; the optimization module is used to optimize the results of the traffic control strategy and perform signal control on the traffic control sub-area based on the optimized traffic control strategy.

[0119] Specifically, the flow prediction module inputs the detected real-time waiting count and the waiting count detected in the previous interval, and segments the waiting count according to set time intervals. In this embodiment, the waiting count includes the sum of the number of vehicles and the number of pedestrians. The flow prediction module consists of three convolutional layers with convolution kernels of different scales. The first convolution layer is used to extract temporal correlations and obtain time series information. The first convolution layer is a 1×1 convolution. The second convolution layer is used to extract spatial dependencies and obtain spatial information. It uses a 3×3 Laplace convolution kernel. The third layer uses a 1×1 convolution kernel to fuse the obtained time series information and spatial information to obtain a flow prediction matrix consistent with the input size.

[0120] The decision module calculates the traffic flow within the time interval based on lane capacity and historical traffic flow, and then calculates the intersection travel cycle based on the traffic flow:

[0121] Z = l k ( D m - L ) / Y ,

[0122] Where, l k Indicates the k Traffic flow in each phase; D m Indicates the optimal cycle; L Indicates the total lost time within the signal period; Y Indicates the maximum traffic flow in all phases.

[0123] An optimization function is constructed for the intersection travel cycle based on nonlinear functions to achieve the optimal configuration of traffic flow and minimize road congestion.

[0124] The optimization function is:

[0125] H = f 1 ( F ,S , E ),

[0126] Where, f 1 is a nonlinear function, such as regression analysis and neural network; F 、 S 、 E They represent traffic status, signal control strategy and environmental factors respectively.

[0127] S4. Build a virtual traffic operation model. Based on the processed data from the test set and the traffic control model, perform a virtual operation of traffic signal control to obtain traffic signal control results. Combined with virtual reality technology, the operational effects of the adopted traffic control strategy can be more intuitively observed.

[0128] Methods for constructing a virtual traffic operation model include:

[0129] Build a dynamic database and a static database; the dynamic database is used to store pedestrian and vehicle models; the static database is used to store road and traffic facility models.

[0130] Using mapping technology, traffic operation scenarios are set based on the historical traffic basic information and processed data of the test set.

[0131] The traffic signal control results are obtained based on the optimized traffic control strategy output by the traffic operation scenario and traffic control model.

[0132] S5. Conduct quantitative evaluation on traffic signal control results and complete traffic signal control.

[0133] In this embodiment, the traffic control results are evaluated by the signal control quantitative evaluation index of each phase; by setting an index threshold, when it is lower than the index threshold, the traffic control method is adjusted. The signal control quantitative evaluation index includes:

[0134] ,

[0135] Where, Score Represents the credit control quantitative evaluation index; BS An equilibrium index indicating the degree of saturation between phases; v Indicates the number of intersection phases; Indicates the signal control quantitative evaluation index of the phase; α represents the weight of pedestrian tolerance in the traffic control quantitative evaluation index; η Indicates pedestrian tolerance.

[0136] in,

[0137] ,

[0138] ,

[0139] Where, ROR 、 、 They represent red light occupancy, queue ratio and green light utilization respectively; 、 、 are the weight coefficients of red light occupancy rate, queue ratio and green light utilization rate respectively; Indicates phase i saturation; Indicates the number of intersection phases.

[0140] Example 2:

[0141] This embodiment provides a traffic signal control optimization system based on big data and deep learning, including: an information collection module, an area division module, a signal control module, and an evaluation module; the information collection module is used to obtain historical traffic basic information and historical traffic flow information within the control area, and pre-process the historical traffic flow information to obtain processed data; wherein, the historical traffic flow information includes video information collected by traffic cameras; the processed data includes: vehicle information and pedestrian information.

[0142] The method for the information collection module to pre-process historical traffic flow information includes:

[0143] Acquire video information, collect frame data of the video information, and obtain a traffic flow image.

[0144] The lane area and pedestrian waiting area of ​​the traffic flow image are divided. Specifically, a lane line detection model is used to detect the lane lines in the traffic flow image, and the lane area is divided based on the detection results.

[0145] In this embodiment, the lane detection model includes an encoder and a decoder. The encoder includes a first convolutional module and a second convolutional module. The first convolutional module includes three first convolutional layers. The second convolutional module includes two second convolutional layers. The first convolutional layer includes 3×3 convolution, batch normalization, and Reluctant Unit (ReLU) activation. A residual block is provided between the two first convolutional layers, using 1×1 convolution. The input or downsampled features are added to the current feature map to achieve efficient residual connections. The second convolutional module includes two branches: a first branch and a second branch. The first and second branches each include two nonlinear activation functions. The nonlinear activation function of the second branch is preceded and followed by two basis functions with a kernel size of 1. The outputs of the two branches are then added. The decoder includes a second convolutional module and a third convolutional module. The third convolutional module includes three third convolutional layers. The third convolutional layer includes 3×3 convolution, batch normalization, and Reluctant Unit (ReLU) activation. An attention connection module is provided between the decoder and encoder. Among them, the attention connection module is improved on the basis of the traditional ELA module. The traditional ELA module uses avg pool in the spatial dimension to obtain the feature vectors in the horizontal and vertical directions. The attention connection module of this embodiment combines avg pool and max pool in the spatial dimension, and max pool is used to obtain the most significant feature values ​​in the local area. The avg pool and max pool are used to obtain the feature vectors in the horizontal and vertical directions respectively, and the obtained feature vectors are spliced. The lane line detection model is optimized by collecting scene images in different situations to obtain the optimal detection model and realize the detection of lane lines. The lane area is divided by the detected lane lines.

[0146] The positions of pedestrians are identified based on the traffic flow image, and pedestrian waiting areas are divided based on the pedestrian positions. In this embodiment, the non-lane area is divided into the pedestrian waiting area.

[0147] Vehicles are identified in the lane area to obtain vehicle information; pedestrians are identified in the pedestrian waiting area to obtain pedestrian information.

[0148] In this embodiment, a vehicle recognition model is used to identify vehicles in a lane area image, and the vehicle's position and average speed over the 10 consecutive seconds before the vehicle stops are determined. The vehicle recognition model includes a vehicle recognition module and a calculation module; the recognition module is used to identify vehicles in the lane area image, and the calculation module is used to calculate speed information for the identified vehicles.

[0149] Specifically, the vehicle recognition module includes a backbone network, a neck network, and a head detection network. The backbone network includes a first backbone module, a second backbone module, a third backbone module, and a fourth backbone module; the neck network includes a first neck module, a second neck module, and a third neck module; and the head detection network includes a detection module. In this embodiment, the first backbone module consists of a convolutional layer (Conv); the input lane area image passes through a 1*1 convolutional layer to extract convolutional features. The second backbone module includes four groups of second backbone units, each group of second backbone units consists of a guide block and an aggregation block; the guide block consists of four feature extraction blocks, namely the first feature extraction block, the second feature extraction block, the third feature extraction block, and the fourth feature extraction block; the obtained convolutional features are respectively input into the first feature extraction block and the second feature extraction block to learn local features and contextual features, thereby obtaining local features and contextual features. In this embodiment, the first feature extraction block consists of a 3*3 standard convolutional layer, and the second feature extraction block consists of a 3*3 depthwise convolution. The resulting local and contextual features are then fed into the third feature extraction module for fusion, yielding the third fused feature. The third feature extraction module consists of a concatenation layer, a batch normalization (BN) layer, and a ReLU activation function. The concatenation layer fuses the local and contextual features, and batch normalization and activation further enhance the feature representation, yielding the third fused feature. The third fused feature is then fed into the fourth feature extraction block, yielding the fourth fused feature. The fourth feature extraction block consists of a global average pooling layer and two fully connected layers. The global average pooling layer aggregates contextual information from the third fused feature, and the feature is then further processed by two fully connected layers to yield the fourth fused feature. The aggregation block consists of two branches: the first and second branches. The fourth fused feature is first channel-compressed by a 1x1 standard convolution and then fed into the first and second branches, respectively. The first branch consists of a reparameterized convolution, a 3x3 standard convolution, and a 1x1 standard convolution. The second branch processes the input using a 1x1 standard convolution. The outputs of the first and second branches are then fed into the concatenation layer for integration, followed by a 1x1 standard convolution. The third backbone module uses a standard spatial pyramid pooling module. The fourth backbone module uses a standard C2PSA module. The neck network includes: a first neck module, a second neck module, and a third neck module. The first neck module consists of a single 1x1 convolutional layer; the second neck module uses the same structure as the second backbone module; and the third neck module uses the same structure as above.The detection module adopts three detection branches, which are composed of P3, P4, and P5 detection layers. The P3, P4, and P5 detection layers are respectively connected to a 1*1 group normalization convolution layer. The group normalization convolution layer groups the feature dimensions of the input image and performs normalization within each group. It then passes through two DE convolution layers to capture more image details. The feature map output by the DE convolution layer then passes through the Conv_Reg and Conv_Cls layers for bounding box regression and category classification, and scale is used to adjust the scale of the final output. In this embodiment, the DE convolution layer includes five convolution layers, namely: ordinary convolution, center difference convolution, angular difference convolution, horizontal difference convolution, and vertical difference convolution. First, the center difference convolution and angular difference convolution are used to encode the prior information into the convolution layer, and then the horizontal difference convolution and vertical difference convolution are used to integrate the traditional local descriptor into the convolution layer, and the representation and generalization capabilities are enhanced by learning gradient information. Normal convolution and differential convolution are used to obtain horizontal intensity and enhance gradient information, and integrate the features learned by the five convolutional layers. Finally, reparameterization technology is used to obtain rich features without introducing additional computational cost, which is used as the output of the DE convolutional layer.

[0150] The calculation module tracks the identified vehicle and obtains the vehicle's speed information by combining the sampling time between the upper and lower frames, the vehicle's movement distance, and the image acquisition ratio. It calculates the average speed of the vehicle over a continuous 10-second period.

[0151] Pedestrian information includes: the number of pedestrians. Pedestrians in the pedestrian waiting area are detected using a neural network model to obtain the number of pedestrians and their intentions. This embodiment uses a pedestrian recognition model to identify the number of pedestrians and their intentions.

[0152] Specifically, the pedestrian recognition model includes: a pedestrian recognition unit and an intention recognition unit; wherein the pedestrian recognition unit adopts the structure of the vehicle recognition unit described above to identify pedestrians. The intention recognition unit is used to track the identified pedestrian and obtain the pedestrian's intention by combining the sampling time between the upper and lower frames and the pedestrian's movement distance. In this embodiment, the pedestrian's intention includes waiting and moving; by setting a motion threshold and comparing it with the obtained pedestrian's movement distance, when the pedestrian's movement distance is greater than the set operating threshold, the pedestrian's intention is judged to be moving, otherwise it is waiting. Pedestrian tolerance is also calculated based on pedestrians with moving intentions and pedestrians with waiting intentions.

[0153] The area division module is used to divide traffic control sub-areas based on basic traffic information.

[0154] The method for dividing traffic control sub-areas includes: extracting road segments and intersections based on historical traffic basic information; calculating the correlation between road segments to obtain a correlation matrix; clustering the road segments based on the correlation matrix to obtain a first set of clusters; calculating correlation coefficients between the first set of clusters; merging the first set of clusters based on the correlation coefficients to obtain a second set of clusters; and completing the division of traffic control sub-areas based on the second set of clusters.

[0155] The method of obtaining the correlation matrix includes:

[0156] A road network model is constructed based on road sections and intersections; the correlation coefficients between intersections are calculated, and a correlation matrix is ​​obtained based on the correlation coefficients and the road network model.

[0157] Incorporating graph theory, the road network in the traffic control area is described as a topological structure graph. The extracted intersections are defined as points, and the road segments connecting them are defined as edges. Each segment is numbered and represented using an undirected graph as an adjacency matrix. In the adjacency matrix, the diagonal elements are set to 0, the matrix elements corresponding to the numbers of adjacent segments are set to 1, and the matrix elements corresponding to non-adjacent segments are set to 0, thus forming a road network model.

[0158] In this embodiment, the calculation method of the correlation coefficient includes:

[0159] ,

[0160] Where, F Q(A,B) represents the correlation coefficient; n 0 indicates the number of traffic branches in the two intersection directions; Q Amax Indicates the maximum flow value of the traffic branch; N Indicates the number of lanes; Q k（A→B） Indicates the k Traffic flow value of the branch road, t AB Indicates that the vehicle is passing through the intersection A , B The average travel time between.

[0161] Methods for obtaining the first cluster include:

[0162] First, randomly divide a traffic control subarea into a number of subareas. Arrange the local density of the first b road segments in descending order, and use the first b road segments as cluster centers. When cluster centers are adjacent in space, the cluster center with the highest local density is selected as the new cluster center. This process is repeated for all cluster centers until all cluster centers are non-adjacent, resulting in a new cluster point set. The shortest path between the remaining road segments and the new cluster point is calculated. Using this shortest path value, the remaining road segments are merged into the new cluster point until clustering is complete, resulting in a number of clusters.

[0163] The calculation method of correlation coefficient includes:

[0164] ,

[0165] ,

[0166] X = RI × βRC ;

[0167] Where, RI represents the first correlation coefficient, RI ( m,n ) represents a cluster m and clusters n The first correlation coefficient between RC represents the second correlation coefficient, RC ( m 1 ,n 1) Represents a subcluster m 1 and subclusters n The second correlation coefficient between 1; X represents the correlation coefficient; EC ( m , n ) represents a connected cluster m and clusters n The sum of the weights of all edges; EC ( m )and EC ( n ) represent clusters m and clusters n The edge weights that are divided into two parts of approximately equal size; Indicates that the cluster C Divide into subclusters m 1 and subclusters n An edge weight of 1; and Respectively represent the sub-clusters m 1 and subclusters n 1 is the edge weight of the partition into two parts of approximately equal size; β Represents the weight of the second correlation coefficient.

[0168] A merging threshold is set, and two clusters with correlation coefficients greater than the merging threshold are merged until all clusters are merged to obtain a second cluster set, which is then divided into corresponding traffic control sub-areas according to the number of clusters in the second cluster set.

[0169] The signal control module is used to build a traffic control model and perform signal control on the traffic control sub-area based on the processed data of the training set and the traffic control model.

[0170] In this embodiment, the traffic control model includes: a traffic prediction module, a decision module and an optimization module; the traffic prediction module is used to predict traffic based on historical traffic, real-time traffic and the next time period; the decision module performs traffic control based on the predicted traffic to obtain a traffic control strategy; the optimization module is used to optimize the results of the traffic control strategy and perform signal control on the traffic control sub-area based on the optimized traffic control strategy.

[0171] Specifically, the flow prediction module inputs the detected real-time waiting count and the waiting count detected in the previous interval, and segments the waiting count according to set time intervals. In this embodiment, the waiting count includes the sum of the number of vehicles and the number of pedestrians. The flow prediction module consists of three convolutional layers with convolution kernels of different scales. The first convolution layer is used to extract temporal correlations and obtain time series information. The first convolution layer is a 1×1 convolution. The second convolution layer is used to extract spatial dependencies and obtain spatial information. It uses a 3×3 Laplace convolution kernel. The third layer uses a 1×1 convolution kernel to fuse the obtained time series information and spatial information to obtain a flow prediction matrix consistent with the input size.

[0172] The decision module calculates the traffic flow within the time interval based on lane capacity and historical traffic flow, and then calculates the intersection travel cycle based on the traffic flow:

[0173] Z = l k ( D m - L ) / Y ,

[0174] Where, l k Indicates the k Traffic flow in each phase; D m Indicates the optimal cycle; L Indicates the total lost time within the signal period; Y Indicates the maximum traffic flow in all phases.

[0175] An optimization function is constructed for the intersection travel cycle based on nonlinear functions to achieve the optimal configuration of traffic flow and minimize road congestion.

[0176] The optimization function is:

[0177] H = f 1 ( F , S , E ),

[0178] Where, f 1 is a nonlinear function, such as regression analysis and neural network; F 、 S 、 E They represent traffic status, signal control strategy and environmental factors respectively.

[0179] The virtual operation module is used to build a virtual traffic operation model. Based on the processed data of the test set and the traffic control model, traffic signal control is simulated and the traffic signal control results are obtained. Combined with virtual reality technology, the operational effects of the adopted traffic control strategy can be more intuitively observed.

[0180] Methods for constructing a virtual traffic operation model include:

[0181] Build a dynamic database and a static database; the dynamic database is used to store pedestrian and vehicle models; the static database is used to store road and traffic facility models.

[0182] Using mapping technology, traffic operation scenarios are set based on the historical traffic basic information and processed data of the test set.

[0183] The evaluation module is used to quantitatively evaluate the control effect of traffic signals and complete the control of traffic signals.

[0184] In this embodiment, the evaluation module evaluates the traffic signal control result by using the signal control quantitative evaluation index of each phase; wherein the signal control quantitative evaluation index includes:

[0185] ,

[0186] Where, Score Represents the credit control quantitative evaluation index; BS An equilibrium index indicating the degree of saturation between phases; v Indicates the number of intersection phases; Indicates the signal control quantitative evaluation index of the phase; α represents the weight of pedestrian tolerance in the traffic control quantitative evaluation index; η Indicates pedestrian tolerance.

[0187] in,

[0188] ,

[0189] ,

[0190] Where, ROR 、 、 They represent red light occupancy, queue ratio and green light utilization respectively; 、 、 are the weight coefficients of red light occupancy rate, queue ratio and green light utilization rate respectively; Indicates phase i saturation; Indicates the number of intersection phases.

[0191] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A traffic signal control optimization method based on big data and deep learning, characterized in that: The method comprises the following steps: S1. Obtain historical basic traffic information and historical traffic flow information within the control area, divide the historical basic traffic information and the historical traffic flow information into a training set and a test set, and preprocess the historical traffic flow information to obtain processed data; S2. Dividing traffic control sub-areas based on the historical traffic basic information; S3. Build a traffic control model, and perform signal control on the traffic control sub-area based on the processed data of the training set and the traffic control model; S4. Construct a virtual traffic operation model, perform a virtual traffic signal control operation based on the processed data of the test set and the traffic control model, and obtain a traffic signal control result; S5. Quantitatively evaluate the traffic signal control result to complete the traffic signal control; The traffic control model includes: a flow prediction module, a decision module and an optimization module; The traffic prediction module is used to predict traffic flow in the next time period based on historical traffic flow and real-time traffic flow; The decision module performs traffic control based on the predicted traffic flow to obtain a traffic control strategy; The optimization module is used to optimize the result of the traffic control strategy and perform signal control on the traffic control sub-area based on the optimized traffic control strategy; The method for constructing the virtual traffic operation model includes: Constructing a dynamic database and a static database; the dynamic database is used to store pedestrian and vehicle models; the static database is used to store road and traffic facility models; Setting a traffic operation scenario based on the historical traffic basic information of the test set and the processed data; Obtaining the traffic signal control result based on the traffic operation scenario and the optimized traffic control strategy output by the traffic control model; The traffic signal control result is quantitatively evaluated using a signal control quantitative evaluation index, wherein the signal control quantitative evaluation index includes: In the formula, Score represents the signal control quantitative evaluation index; BS represents the saturation balance index between phases; v represents the number of intersection phases; Score phase represents the phase signal control quantitative evaluation index; α represents the weight of pedestrian tolerance in the signal control quantitative evaluation index; η represents pedestrian tolerance; in, Score phase =(β1ROR+β2R p +β3R g )×100 Where ROR, R P 、R g They represent the red light occupancy rate, queue ratio and green light utilization rate respectively; β1, β2 and β3 are the weight coefficients of the red light occupancy rate, queue ratio and green light utilization rate respectively; s i represents the saturation of phase i; n2 represents the number of intersection phases.

2. The traffic signal control optimization method based on big data and deep learning according to claim 1 is characterized in that: In S1, the historical traffic flow information includes video information collected by traffic cameras; The processed data includes: vehicle information and pedestrian information; The method for obtaining the processed data includes: S12, acquiring the video information, collecting frame data of the video information, and obtaining a traffic flow image; S13, dividing the traffic flow image into a lane area and a pedestrian waiting area; S14, performing vehicle identification on the lane area to obtain vehicle information; S15. Perform pedestrian recognition in the pedestrian waiting area to obtain pedestrian information.

3. The traffic signal control optimization method based on big data and deep learning according to claim 2 is characterized in that: In S13, the method of dividing the lane area and the pedestrian waiting area includes: Detecting lane lines in the traffic flow image using a lane line detection model, and dividing the lane area based on the detection results; The positions of pedestrians are identified based on the traffic flow image, and the pedestrian waiting area is divided based on the positions of pedestrians.

4. The traffic signal control optimization method based on big data and deep learning according to claim 3 is characterized in that: The lane line detection model includes: an encoder part and a decoder part; The encoder part includes: a first convolution module and a second convolution module; the first convolution module includes three first convolution layers; the second convolution module includes two second convolution layers; The first convolutional layer includes: 3×3 convolution, batch normalization and ReLU activation; a residual block is set between the two first convolutional layers, and the residual block adopts 1×1 convolution; The second convolution module includes two branches: a first branch and a second branch, each of the first branch and the second branch includes two nonlinear activation functions, and the nonlinear activation function of the second branch is preceded and followed by two basis functions with a kernel size of 1; the outputs of the two branches are then added; The decoder part includes: the second convolution module and the third convolution module; The third convolution module includes three third convolution layers; the third convolution layers include: 3×3 convolution, batch normalization and ReLU activation; An attention connection module is provided between the decoder part and the encoder part.

5. The traffic signal control optimization method based on big data and deep learning according to claim 1 is characterized in that: In S2, the method of dividing the traffic control sub-area includes: S21, extracting road sections and intersections based on the historical traffic basic information; S22, calculating the correlation between the road sections to obtain a correlation matrix; S23, clustering the road sections based on the association matrix to obtain a first cluster set; S24. Calculate the correlation coefficient between the first clusters, merge the first clusters based on the correlation coefficient to obtain a second cluster; and complete the division of the traffic control sub-area based on the second cluster. The calculation method of the correlation coefficient includes: X = RI × βRC; Where RI represents the first correlation coefficient, RI(m,n) represents the first correlation coefficient between cluster m and cluster n; RC represents the second correlation coefficient, RC(m1,n1) represents the second correlation coefficient between sub-cluster m1 and sub-cluster n1; X represents the correlation coefficient; EC(m,n) represents the sum of the weights of all edges connecting cluster m and cluster n; EC(m) and EC(n) represent the edge weights of dividing cluster m and cluster n into two parts of approximately equal size, respectively. represents the edge weight that divides cluster C into sub-clusters m1 and n1; and They represent the edge weights for dividing sub-cluster m1 and sub-cluster n1 into two parts of approximately equal size; β represents the weight of the second correlation coefficient.

6. The traffic signal control optimization method based on big data and deep learning according to claim 5 is characterized in that: The method for obtaining the correlation matrix includes: Constructing a road network model according to the road sections and the intersections; Calculating correlation coefficients between the intersections, and obtaining the correlation matrix based on the correlation coefficients and the road network model; The calculation method of the correlation coefficient includes: Where, F Q(A,B) represents the correlation coefficient; n0 represents the number of traffic branches in the two intersection directions; Q Amax represents the maximum flow value of the traffic branch; N represents the number of lanes; Q k(A→B) represents the traffic flow value of the kth branch, t AB It represents the average travel time of vehicles passing through intersections A and B.

7. A traffic signal control optimization system based on big data and deep learning, wherein the control optimization system is used to implement the method according to any one of claims 1 to 6, characterized in that: include: Information collection module, area division module, signal control module, virtual operation module and evaluation module; The information collection module is used to obtain historical traffic basic information and historical traffic flow information in the control area, divide the historical traffic basic information and the historical traffic flow information into training sets and test sets, and pre-process the historical traffic flow information to obtain processed data; The area division module is used to divide the traffic control sub-areas based on the historical traffic basic information; The signal control module is used to construct a traffic control model and perform signal control on the traffic control sub-area based on the processed data of the training set and the traffic control model; The virtual operation module is used to build a virtual traffic operation model, perform a virtual operation of traffic signal control based on the processed data of the test set and the traffic control model, and obtain a traffic signal control result; The evaluation module is used to perform quantitative evaluation on the traffic signal control result to complete the control of the traffic signal.

Images