A Smart Transportation Data Management and Control Method and System Based on Edge Computing

Abstract

This invention discloses a smart traffic data management and control method and system based on edge computing, comprising the following steps: S1, collecting multi-dimensional traffic perception data at intersections and edge node resource status parameters; S2, segmenting the foreground of the video stream and mapping the radar point cloud to a unified coordinate system to generate a standardized edge input dataset; S3, generating a dynamic management and control weight set using a multi-view spatiotemporal interactive attention network; S4, extracting traffic situation features based on the weight set and generating queue instructions; S5, mapping the features to a local microscopic digital twin and generating a handover data packet; S6, sending the handover data packet and introducing a bilinear pooling mechanism using an improved squeezing and excitation residual network to trigger precise evaporation of non-critical data; S7, reading the status and encapsulating the output management and control data packet. This invention achieves efficient diversion and precise evaporation of traffic data, alleviates the storage pressure on edge nodes, and improves real-time transmission performance.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and edge computing, and in particular to an intelligent transportation data management and control method and system based on edge computing. Background Technology

[0002] Edge computing technology, with its advantages of low latency, fast response, and reduced cloud load, has been widely used in recent years in fields such as smart transportation, intelligent connected vehicles, and autonomous driving, becoming an important architectural support for the next generation of traffic management systems. However, in practical applications, traffic intersection data management scenarios face many challenges, such as massive heterogeneous sensing data, limited network bandwidth resources, and fluctuating computing capabilities of edge nodes. The effectiveness of data management based on edge computing is still constrained by many factors.

[0003] Most current traffic data management methods rely on single-modal data analysis, making it difficult to fully utilize multi-source heterogeneous information such as vehicle video streams, radar point cloud data, and roadside unit information. This results in a lack of comprehensiveness and accuracy in traffic situation modeling. Some systems only use fixed thresholds or static rules for data distribution, ignoring the combined impact of multiple factors such as real-time computing resource status, data timeliness sensitivity, and scenario complexity, thus limiting the adaptive adjustment capability of management strategies. Furthermore, the data filtering and transmission logic lacks intelligent feedback mechanisms, making it difficult to dynamically adjust computing resource allocation strategies under complex road conditions. This can lead to critical data being discarded or delayed due to resource contention, affecting the effectiveness and real-time nature of management results.

[0004] Furthermore, existing edge control systems often employ rigid mechanisms to handle resource overload, failing to accurately assess the value of non-critical data and handle data loss based on dynamic changes in load status. This makes the system highly susceptible to computational blockages and even service crashes under high-concurrency scenarios. Simultaneously, the lack of feature extraction methods combining bilinear pooling and residual structures prevents the capture of high-order interaction information between resource status feature channels, hindering the generation of robust weight recalibration parameters and severely impacting the system's stability and reliability in real-world traffic scenarios. Summary of the Invention

[0005] One objective of this invention is to propose a smart traffic data management and control method and system based on edge computing. This invention fully integrates key steps such as multi-source perception spatiotemporal alignment, multi-view spatiotemporal interactive attention network evaluation, improved squeezing and incentive residual network feature extraction, and precise data evaporation. It constructs a traffic intelligent management and control process with standardized data input, dynamic management weight generation, second-order interaction capture of feature channels, and simplified data stream output, achieving comprehensive perception and adaptive load optimization of traffic data at resource-constrained edge nodes. By introducing a bilinear pooling mechanism into the residual branch, this invention effectively captures second-order interaction information of feature channels, possessing advantages such as high accuracy in data value assessment, strong robustness in resource state feature extraction, and flexible response to load overload. It can significantly improve the data processing efficiency and system stability at the edge, thereby effectively solving problems such as insufficient resource feature expression, rigid overload protection mechanisms, and easy loss of high-value data in existing technologies.

[0006] A smart transportation data management and control method based on edge computing according to an embodiment of the present invention includes the following steps:

[0007] S1. Synchronously collect multi-dimensional traffic perception data at intersections and obtain real-time computing resource status parameters of the current edge computing node;

[0008] S2. Perform real-time background modeling and foreground segmentation on vehicle video streams in multi-dimensional traffic perception data, extract pixel-level masks and visual features, map radar point cloud data to a unified geospatial coordinate system, and align it with vehicle visual features in time and space to generate a standardized edge input dataset.

[0009] S3. Input the standardized edge input dataset into the edge computing node, use the multi-view spatiotemporal interactive attention network to extract potential value representations, calculate the data entropy value and time sensitivity of the current traffic scenario, and combine real-time computing resource status parameters to generate a dynamic control weight set.

[0010] S4. Based on the dynamic control weight set, implement adaptive traffic splitting processing on the standardized edge input dataset. For data with high weight and containing complex road condition features, perform real-time structured processing and extract traffic situation features at the local edge node. For low-weight regular traffic flow data, compress and encapsulate it and generate queue instructions.

[0011] S5. Map traffic situation features to a local micro-digital twin constructed inside the edge node, and predict the coverage area of ​​the next-hop edge node based on the vehicle movement trajectory to generate a handover data packet.

[0012] S6. The handover data packet is sent to the predicted next-hop edge node. The response queue command uploads regular traffic flow data to the cloud center. An improved squeezing and excitation residual network is used for feature extraction. A bilinear pooling mechanism is introduced into the residual branch to capture second-order interaction information of feature channels, generating weight recalibration parameters. When the load exceeds limits, a precise evaporation operation of non-critical data is triggered, generating a simplified data stream. The real-time output monitors the operating status of the local microscopic digital twin.

[0013] S7. Read the running status and simplified data stream, and concatenate them according to the preset data structure order. Then, encapsulate them into a control data packet of preset format and output it.

[0014] Optionally, S1 specifically includes: synchronously collecting multi-dimensional traffic perception data through sensing devices deployed at the intersection, the multi-dimensional traffic perception data including vehicle video streams, radar point cloud data, and vehicle speed and location information uploaded by roadside units; calling the system monitoring interface of the edge computing node to read the current operating indicators; and obtaining the real-time computing resource status parameters of the edge computing node, the real-time computing resource status parameters including CPU utilization, remaining memory, and network bandwidth load.

[0015] Optionally, S2 specifically includes:

[0016] S21. Read the vehicle video stream from the multidimensional traffic perception data, use the Gaussian mixture model to statistically model the pixels in the continuous video frames, calculate the difference between the pixel value of the current frame and the mean of the background model, determine the pixel area with a difference greater than the preset difference threshold as the foreground area, and extract the pixel-level mask of the vehicle target.

[0017] S22. Perform a convolution operation of a preset size on the image area covered by the pixel-level mask, calculate the gradient magnitude and direction values ​​pixel by pixel, and generate visual features.

[0018] S23. Acquire radar point cloud data. Based on the image features and point cloud coordinates of the calibration board in the public field of view collected in advance, use the PnP algorithm to solve the spatial transformation relationship between the radar coordinate system and the camera coordinate system. Then, calculate the rotation matrix and translation vector by minimizing the reprojection error through nonlinear optimization. Map the radar point cloud data to a unified geospatial coordinate system so that the point cloud coordinates and the pixel coordinates in the vehicle video stream are established in correspondence.

[0019] S24. Based on the correspondence, the radar point cloud data is projected onto the image plane where the visual features are located, and the times of the two are aligned according to the timestamps of the data acquisition to generate a standardized edge input dataset.

[0020] Optionally, S3 specifically includes:

[0021] S31. Input the standardized edge input dataset into the edge computing node to construct a multi-view spatiotemporal interactive attention network. Preset the first feature encoding branch and the second state encoding branch for parallel input. Input the standardized edge input dataset into the first feature encoding branch and the real-time computing resource state parameters into the second state encoding branch. Perform tensor concatenation operation on the outputs of the two branches.

[0022] S32. Inside the first feature encoding branch, a one-dimensional temporal convolution kernel of a preset size is set, and a sliding window is performed on the standardized edge input dataset along the time axis. The dot product value of the feature vectors in the window is calculated and scaled by dividing by the square root of the vector dimension. It is then converted into a probability value through the Softmax function to generate a temporal dependent weight matrix and multiplied with the standardized edge input dataset to extract data content features.

[0023] S33. Inside the second state coding branch, a multi-layer perceptron structure containing a preset fully connected layer is constructed. The CPU utilization, memory remaining and network bandwidth load in the real-time computing resource state parameters are multiplied by a preset weight matrix and superimposed with a bias vector. After the first linear transformation, the ReLU activation function is input for nonlinear mapping and then after the second linear transformation, the edge resource state features are extracted.

[0024] S34. Using data content features as query vectors and edge resource status features as key and value vectors, calculate the product of the query vector and the transpose of the key vector to obtain the attention score matrix, and input it into the Softmax function for normalization. Multiply the result with the value vector matrix to establish a global association between data content features and edge resource status features, and generate a potential value representation.

[0025] S35. Perform statistical distribution analysis on the potential value representation, count the number of samples of different data categories in the traffic scenario, divide by the total number of samples to calculate the probability of occurrence of each category, multiply the probability of occurrence by the natural logarithm function and sum over all categories to obtain the data entropy value, and at the same time read the generated timestamp, calculate the difference between the current timestamp and the generated timestamp, divide the difference by the preset time decay constant to obtain the timeliness sensitivity.

[0026] S36. Construct a weight generation network containing two preset fully connected layers. Concatenate the data entropy value, time sensitivity, and real-time computing resource status parameters on the feature dimension, input them into the first fully connected layer, multiply them by preset weights and add biases, process them through the Tanh activation function and input them into the second fully connected layer, and map the output result to the zero to one interval through the Sigmoid activation function to generate a dynamically controlled weight set.

[0027] Optionally, S4 specifically includes:

[0028] S41. Read the weight value corresponding to each standardized edge input dataset in the dynamic control weight set, compare the weight value with the preset high weight diversion threshold. If the weight value is greater than or equal to the high weight diversion threshold, it is determined to be high weight data; otherwise, it is determined to be low weight data.

[0029] S42. For high-weight data, read vehicle video stream images from the standardized edge input dataset, perform convolution operations using a pre-defined convolutional neural network, extract feature maps, and generate vehicle bounding box coordinates and vehicle type labels through regression calculation.

[0030] S43. Simultaneously read radar point cloud data, calculate the Euclidean distance between point clouds, group points with a distance less than the preset clustering threshold into one class, calculate the position change of the cluster center point, divide by the corresponding time interval to obtain the real-time speed and heading angle of the vehicle, and splice the vehicle bounding box coordinates, vehicle type label, real-time speed and heading angle of the vehicle according to the preset data structure order to extract traffic situation features.

[0031] S44. For low-weight data, read the vehicle video stream from the standardized edge input dataset, use the H.264 coding standard to perform discrete cosine transform on the image frames in the video stream, quantize the transformed frequency coefficients, discard high-frequency coefficients, and retain low-frequency coefficients.

[0032] S45. Simultaneously read radar point cloud data, preset the size of the voxel grid, divide the point cloud space into multiple cubic grids, calculate the average coordinates of the point cloud in each grid, replace all the original point cloud coordinates in the grid with the average coordinates, package and encapsulate the processed video stream data and point cloud data in binary data format, write the target storage address in the header of the encapsulated data, and generate a queue instruction to upload to the cloud center.

[0033] Optionally, S5 specifically includes:

[0034] S51. Create a 3D virtual scene model containing road geometric topology information as a local micro digital twin, read traffic situation characteristics, map vehicle bounding box coordinates and vehicle type labels to the corresponding positions in the 3D virtual scene model, and generate a virtual vehicle model consistent with the real-time traffic state.

[0035] S52. In the three-dimensional virtual scene model, read the position coordinates of the same vehicle at consecutive times in the order of timestamps, calculate the Euclidean distance vector between the position coordinates at adjacent times, and connect all the Euclidean distance vectors end to end to form the continuous movement trajectory curve of the vehicle.

[0036] S53. Read the coordinates of the end point of the continuous movement trajectory curve as the starting point, obtain the instantaneous movement direction of the vehicle at the starting point, construct a ray of preset length, calculate the intersection point of the ray with the boundary of the preset circular coverage area of ​​all edge nodes in the three-dimensional virtual scene model, filter out the intersection point closest to the starting point, and determine the circular coverage area where the intersection point is located as the coverage area of ​​the next hop edge node that the vehicle is about to enter.

[0037] S54. Extract the real-time speed and heading angle of vehicles from traffic situation characteristics, read the vehicle identification, package the vehicle identification, continuous movement trajectory curve, real-time speed and heading angle of vehicles, and current state attributes of virtual vehicle models into a data package to generate a handover data package.

[0038] Optionally, S6 specifically includes:

[0039] S61. Establish communication links between edge nodes, read the vehicle identification and the predicted coverage area of ​​the next-hop edge node from the handover data packet, call the preset cooperative communication protocol interface, encapsulate the handover data packet into a transmission data frame, send it to the next-hop edge node, and at the same time read the queue instruction uploaded to the cloud center, parse the target storage address in the instruction, read the regular traffic flow data into a binary stream, and send it to the cloud center through the transmission control protocol.

[0040] S62. Read the real-time computing resource status parameters of the edge computing node, and concatenate the CPU utilization, memory remaining amount and network bandwidth load into a three-dimensional resource status vector in numerical order. Input the three-dimensional resource status vector into the input layer of the improved squeezing and excitation residual network, and use a convolution kernel of preset size to perform sliding window convolution operation on the three-dimensional resource status vector along the time step to extract the preliminary resource feature map.

[0041] S63. Construct a residual main branch within the improved squeeze and excitation residual network. Input the preliminary resource feature map into the first convolutional unit. Set the first convolutional unit to contain 64 convolutional kernels. Perform dimensionality-up convolution on the preliminary resource feature map using a convolution operation with a preset stride of 1. Input the output features into the second convolutional unit. Set the second convolutional unit to contain 32 convolutional kernels. Perform dimensionality-down convolution on the features using a convolution operation with a preset stride of 1. Generate a residual feature map and add it element-wise to the preliminary resource feature map to obtain a fused feature map.

[0042] S64. After the residual main branch of the improved squeezing and excitation residual network, set a bilinear pooling branch in parallel, read the fused feature map and make a copy to obtain a copy feature map, calculate the product of the transpose matrix of the fused feature map and the copy feature map, normalize each value in the product result matrix by dividing it by the total number of spatial pixels of the fused feature map, and generate a second-order interactive feature matrix.

[0043] S65. Perform global average pooling on the second-order interaction feature matrix along the spatial width and height dimensions to compress the two-dimensional matrix into a one-dimensional channel description vector and input it into the first fully connected layer. The preset number of neurons is one-quarter of the number of channels. Perform linear dimensionality reduction on the one-dimensional channel description vector, perform ReLU activation, and input it into the second fully connected layer to restore the dimension to the original number of channels and generate weight recalibration parameters through the Sigmoid activation function.

[0044] S66. Multiply the weight recalibration parameters with the fused feature map according to the corresponding positions of each channel, recalibrate the weight of each channel of the fused feature map, read the preset resource load over-limit threshold, calculate the difference between the current CPU utilization and the resource load over-limit threshold, if the difference is greater than zero, determine that the load is over-limit, and trigger the precise evaporation operation of non-critical data.

[0045] S67. Respond to the precise evaporation operation of non-critical data, traverse the standardized edge input dataset in the local cache, read the dynamic control weight set value corresponding to each handover data packet, preset the data evaporation threshold, and filter out data packets with dynamic control weight values ​​less than the data evaporation threshold.

[0046] S68. Mark the selected data packets as non-critical data, delete the original vehicle video stream and radar point cloud data in the non-critical data packets, and extract and retain only the vehicle bounding box coordinates, vehicle type labels, vehicle real-time speed and heading angle. Encapsulate the extracted data into a simplified data stream that retains only structured feature data.

[0047] S69. Read the virtual vehicle model in the local micro digital twin, read the three-dimensional position coordinates of the virtual vehicle model according to the preset sampling interval, calculate the difference of the three-dimensional position coordinates at adjacent sampling times, count the queue length values ​​of all virtual vehicle models, combine the three-dimensional position coordinate differences with the queue length values, and output the current running status.

[0048] Optionally, S7 specifically includes:

[0049] S71. Read the running status and simplified data stream of the local microscopic digital twin, extract the lane queue length value contained in the running status, and extract the vehicle bounding box coordinates, vehicle type label, vehicle real-time speed and heading angle contained in the simplified data stream. Then, concatenate the lane queue length value and the extracted vehicle feature data according to the preset data structure order.

[0050] S72. Encapsulate the spliced ​​data into a control data packet of a preset format, and send the control data packet to the data receiving interface of the traffic control equipment at the intersection to complete the closed-loop control of traffic data on the edge side.

[0051] According to an embodiment of the present invention, a smart transportation data management and control system based on edge computing includes:

[0052] Multi-dimensional traffic data synchronous acquisition and resource perception module: used to synchronously acquire multi-dimensional traffic perception data at intersections and obtain real-time computing resource status parameters of the current edge computing node;

[0053] Multi-source data standardization and fusion module: used to perform real-time background modeling and foreground segmentation on vehicle video streams in multi-dimensional traffic perception data, map radar point cloud data to a unified geospatial coordinate system, and align it with vehicle visual features in time and space to generate a standardized edge input dataset.

[0054] Dynamic weight generation module: It is used to input the standardized edge input dataset into the edge computing node, extract potential value representations by using a multi-view spatiotemporal interactive attention network, and generate a dynamic control weight set by calculating the data entropy value and time sensitivity of the current traffic scenario and combining it with real-time computing resource status parameters.

[0055] Adaptive traffic splitting module: It is used to implement adaptive traffic splitting on the standardized edge input dataset according to the dynamic control weight set. For data with high weight and containing complex road condition features, it performs real-time structured processing and extracts traffic situation features at the local edge node. For low-weight regular traffic flow data, it compresses and encapsulates the data and generates queue instructions.

[0056] Digital twin mapping and handover processing module: used to map traffic situation features to the local micro-digital twins constructed inside the edge nodes, and predict the coverage area of ​​the next-hop edge node based on the vehicle movement trajectory, and generate handover data packets;

[0057] The Deep Feature Extraction and Precise Evaporation Control Module is used to send the handover data packet to the predicted next-hop edge node, respond to queue instructions to upload regular traffic flow data to the cloud center, and use an improved squeezing and excitation residual network for feature extraction. A bilinear pooling mechanism is introduced in the residual branch to capture the second-order interaction information of the feature channel, generate weight recalibration parameters, trigger the precise evaporation operation of non-critical data when the load exceeds the limit, generate a simplified data stream, and output the real-time monitoring status of the local micro digital twin.

[0058] Control data output module: Used to read the running status and simplified data stream, splice them in the order of the preset data structure, encapsulate them into a control data packet of preset format and output them.

[0059] The beneficial effects of this invention are:

[0060] This invention addresses the issues of heterogeneous data sources and inconsistent spatiotemporal references in edge computing environments by establishing a unified mapping between multidimensional traffic perception data and a geospatial coordinate system. It employs the PnP algorithm and nonlinear optimization to perform coordinate transformation and timestamp alignment, generating a standardized edge input dataset. A multi-view spatiotemporal interactive attention network is used to combine data content features with edge resource status features to perform tensor splicing and global correlation analysis, generating a dynamic control weight set. During data processing, high-weight traffic situation features are extracted and low-weight data is compressed and uploaded based on weight thresholds. When resources exceed limits, an improved squeezing and incentivized residual network using a bilinear pooling mechanism is introduced to capture second-order interactive information from resource feature channels. Combined with weight recalibration parameters, non-critical data is precisely evaporated, generating a simplified data stream that retains only structured features. Finally, the local microscopic digital twin's operating status and the simplified data stream are combined for data splicing and encapsulation, and then distributed to intersection control equipment. It enables intelligent management and control of edge data in smart transportation, including full-link value assessment, adaptive adjustment of resource load, and closed-loop output of core control information, effectively improving the utilization rate of edge computing resources, the timeliness of data transmission, and the robustness of the system under complex and high-load scenarios. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Figure 1 This is a flowchart of a smart transportation data management and control method based on edge computing proposed in this invention;

[0063] Figure 2 This is a structural diagram of an intelligent transportation data management and control system based on edge computing proposed in this invention;

[0064] Figure 3 This is a flowchart of the improved squeezing and excitation residual network feature extraction and precise evaporation of non-critical data based on the bilinear pooling mechanism proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figures 1-3 A smart transportation data management and control method based on edge computing includes the following steps:

[0067] S1. Synchronously collect multi-dimensional traffic perception data at the intersection, including vehicle video streams, radar point cloud data, and vehicle speed and location information uploaded by roadside units, and obtain real-time computing resource status parameters of the current edge computing node, including CPU utilization, remaining memory, and network bandwidth load.

[0068] S2. Perform real-time background modeling and foreground segmentation on vehicle video streams in multi-dimensional traffic perception data, extract pixel-level masks and visual features, map radar point cloud data to a unified geospatial coordinate system, and align it with vehicle visual features in time and space to generate a standardized edge input dataset containing multimodal perception information.

[0069] S3. Input the standardized edge input dataset into the edge computing node, use the multi-view spatiotemporal interactive attention network to extract potential value representations, calculate the data entropy value and time sensitivity of the current traffic scenario, and combine real-time computing resource status parameters to generate a dynamic control weight set.

[0070] S4. Based on the dynamic control weight set, implement adaptive traffic splitting processing on the standardized edge input dataset. For data with high weight and containing complex road condition features, perform real-time structured processing and extract traffic situation features at local edge nodes. For low-weight regular traffic flow data, compress and encapsulate it and generate queue instructions to be uploaded to the cloud center.

[0071] S5. Construct a local micro-digital twin within the edge node, map traffic situation features into the local micro-digital twin, and predict the coverage area of ​​the next-hop edge node based on the vehicle movement trajectory, generating a handover data packet containing vehicle identification, historical trajectory features and current digital twin status.

[0072] S6. Based on the collaborative communication protocol between edge nodes, the handover data packet is sent to the predicted next-hop edge node. In response to the queue command to upload to the cloud center, regular traffic flow data is uploaded to the cloud center. An improved squeezing and incentive residual network is used to extract features from real-time computing resource state parameters. A bilinear pooling mechanism is introduced into the residual branch to capture second-order interaction information of feature channels, generating highly robust weight recalibration parameters. When the load exceeds limits, a precise evaporation operation of non-critical data is triggered, generating a simplified data stream that retains only structured feature data. Local micro-digital twins are monitored in real time, and the current operating status is output.

[0073] S7. Read the running status and simplified data stream, and concatenate them according to the preset data structure order. Then, encapsulate them into a control data packet of preset format and output it.

[0074] This invention significantly improves the efficiency and real-time performance of intelligent transportation data management. By synchronously collecting multi-dimensional traffic perception data at intersections and acquiring edge node resource parameters, it achieves spatiotemporal alignment and standardized processing of multi-source heterogeneous data. Utilizing a multi-view spatiotemporal interactive attention network to extract potential value representations, and combining data entropy, time sensitivity, and resource status to generate a dynamic management weight set, it implements precise adaptive traffic diversion processing, ensuring that high-value, complex road condition features are prioritized. Constructing a local microscopic digital twin and predicting the next-hop coverage area based on vehicle trajectories generates handover data packets, enhancing inter-node collaboration. Introducing an improved squeezing and excitation residual network, it uses a bilinear pooling mechanism in the residual branches to capture second-order interaction information of feature channels, generating highly robust weight recalibration parameters. When the load exceeds limits, it triggers precise evaporation of non-critical data, retaining only a streamlined data stream. This method effectively alleviates the storage pressure and network bandwidth bottleneck of edge nodes, significantly reduces transmission latency, ensures the integrity of traffic situation characteristics and real-time monitoring of the digital twin's operating status, and significantly improves the resource utilization and intelligent management level of the intelligent transportation system.

[0075] In this embodiment, S1 specifically includes: synchronously collecting multi-dimensional traffic perception data through sensing devices deployed at the intersection, the multi-dimensional traffic perception data including vehicle video streams, radar point cloud data, and vehicle speed and location information uploaded by roadside units; calling the system monitoring interface of the edge computing node to read the current operating indicators; and obtaining the real-time computing resource status parameters of the edge computing node, the real-time computing resource status parameters including CPU utilization, remaining memory, and network bandwidth load.

[0076] In this embodiment, S2 specifically includes:

[0077] S21. Read the vehicle video stream from the multidimensional traffic perception data, use the Gaussian mixture model to statistically model the pixels in the continuous video frames, calculate the difference between the pixel value of the current frame and the mean of the background model, determine the pixel area with a difference greater than the preset difference threshold as the foreground area, and extract the pixel-level mask of the vehicle target.

[0078] S22. Perform a convolution operation of a preset size on the image area covered by the pixel-level mask, calculate the gradient magnitude and direction values ​​pixel by pixel, and generate visual features containing vehicle texture and shape information.

[0079] S23. Acquire radar point cloud data. Based on the image features and point cloud coordinates of the calibration board in the public field of view collected in advance, use the PnP algorithm to solve the spatial transformation relationship between the radar coordinate system and the camera coordinate system. Then, calculate the rotation matrix and translation vector by minimizing the reprojection error through nonlinear optimization. Map the radar point cloud data to a unified geospatial coordinate system so that the point cloud coordinates and the pixel coordinates in the vehicle video stream are established in correspondence.

[0080] S24. Based on the correspondence, project the radar point cloud data onto the image plane where the visual features are located, and align the times of the two data acquisitions according to the timestamps to generate a standardized edge input dataset.

[0081] In this embodiment, S3 specifically includes:

[0082] S31. Input the standardized edge input dataset into the edge computing node to construct a multi-view spatiotemporal interactive attention network. Preset the first feature encoding branch and the second state encoding branch for parallel input. Input the standardized edge input dataset into the first feature encoding branch and the real-time computing resource state parameters into the second state encoding branch. Perform tensor concatenation operation on the outputs of the two branches.

[0083] S32. Inside the first feature encoding branch, a one-dimensional temporal convolution kernel of a preset size is set. The one-dimensional temporal convolution kernel is used to perform a sliding window scan on the input standardized edge input dataset along the time axis. The dot product value of the feature vectors in the window is calculated and divided by the square root of the vector dimension for scaling. It is then converted into a probability value through the Softmax function to generate a temporal dependent weight matrix and multiplied with the standardized edge input dataset to extract data content features containing the time evolution law.

[0084] S33. Inside the second state coding branch, a multi-layer perceptron structure containing a preset fully connected layer is constructed. The CPU utilization, memory remaining and network bandwidth load in the real-time computing resource state parameters are multiplied by a preset weight matrix and superimposed with a bias vector. After the first linear transformation, the input is given to the ReLU activation function for nonlinear mapping and then after the second linear transformation, the edge resource state features representing the current computing load intensity are extracted.

[0085] S34. Using data content features as query vectors and edge resource status features as key and value vectors, calculate the product of the query vector and the transpose of the key vector to obtain the attention score matrix, and input it into the Softmax function for normalization. Multiply the result with the value vector matrix to establish a global association between data content features and edge resource status features, and generate a potential value representation containing cross-modal information.

[0086] S35. Perform statistical distribution analysis on the potential value representation, count the number of samples of different data categories in the traffic scenario, divide by the total number of samples to calculate the probability of occurrence of each category, multiply the probability of occurrence by the natural logarithm function and sum over all categories to obtain the data entropy value, and at the same time read the generated timestamp, calculate the difference between the current timestamp and the generated timestamp, divide the difference by the preset time decay constant to obtain the timeliness sensitivity.

[0087] S36. Construct a weight generation network containing two preset fully connected layers. Concatenate the data entropy value, time sensitivity, and real-time computing resource status parameters on the feature dimension, input them into the first fully connected layer, multiply them by preset weights and add biases, process them through the Tanh activation function and input them into the second fully connected layer, and map the output result to the zero to one interval through the Sigmoid activation function to generate a dynamically controlled weight set.

[0088] This implementation method constructs a multi-view spatiotemporal interactive attention network as its core innovative technology, which has significant differences and advantages compared to traditional single-modal data transmission or static threshold scheduling methods. Traditional methods often process data content and resource status separately, or only rely on manually set fixed rules for traffic diversion, making it difficult to adapt to dynamically changing traffic scenarios and unable to accurately extract data value in complex environments, resulting in delays or waste of resources in critical data transmission.

[0089] This invention extracts data content features containing temporal evolution patterns and edge resource state features representing computational load intensity by using a first feature encoding branch and a second state encoding branch with pre-defined parallel inputs, respectively, through one-dimensional temporal convolution and a multilayer perceptron structure. In particular, the attention mechanism introduced in S34 establishes a global association between the data content features as query vectors and the resource state features as key-value pairs, generating a potential value representation containing cross-modal information. Based on this, a weight generation network is constructed to output a dynamic control weight set, combining data entropy and time sensitivity. This mechanism enables flexible adjustment of data processing priorities in resource-constrained edge nodes according to real-time scenario requirements, significantly improving the utilization efficiency and control accuracy of traffic perception data under complex road conditions, and effectively solving the problem of high-concurrency data processing at the edge.

[0090] In this embodiment, S4 specifically includes:

[0091] S41. Read the weight value corresponding to each standardized edge input dataset in the dynamic control weight set, compare the weight value with the preset high weight diversion threshold. If the weight value is greater than or equal to the high weight diversion threshold, it is determined to be high weight data; otherwise, it is determined to be low weight data.

[0092] S42. For high-weight data, read vehicle video stream images from the standardized edge input dataset, perform convolution operations using a pre-defined convolutional neural network, extract feature maps, and generate vehicle bounding box coordinates and vehicle type labels through regression calculation.

[0093] S43. Simultaneously read radar point cloud data, calculate the Euclidean distance between point clouds, group points with a distance less than the preset clustering threshold into one class, calculate the position change of the cluster center point, divide by the corresponding time interval to obtain the real-time speed and heading angle of the vehicle, and splice the vehicle bounding box coordinates, vehicle type label, real-time speed and heading angle of the vehicle according to the preset data structure order to extract traffic situation features.

[0094] S44. For low-weight data, read the vehicle video stream from the standardized edge input dataset, use the H.264 coding standard to perform discrete cosine transform on the image frames in the video stream, quantize the transformed frequency coefficients, discard high-frequency coefficients, and retain low-frequency coefficients.

[0095] S45. Simultaneously read radar point cloud data, preset the size of the voxel grid, divide the point cloud space into multiple cubic grids, calculate the average coordinates of the point cloud in each grid, replace all the original point cloud coordinates in the grid with the average coordinates, package and encapsulate the processed video stream data and point cloud data in binary data format, write the target storage address in the header of the encapsulated data, and generate a queue instruction to upload to the cloud center.

[0096] In this embodiment, S5 specifically includes:

[0097] S51. Create a 3D virtual scene model containing road geometric topology information as a local micro digital twin, read traffic situation characteristics, map vehicle bounding box coordinates and vehicle type labels to the corresponding positions in the 3D virtual scene model, and generate a virtual vehicle model consistent with the real-time traffic state.

[0098] S52. In the three-dimensional virtual scene model, read the position coordinates of the same vehicle at consecutive times in the order of timestamps, calculate the Euclidean distance vector between the position coordinates at adjacent times, and connect all the Euclidean distance vectors end to end to form the continuous movement trajectory curve of the vehicle.

[0099] S53. Read the coordinates of the end point of the continuous movement trajectory curve as the starting point, obtain the instantaneous movement direction of the vehicle at the starting point, construct a ray of preset length, calculate the intersection point of the ray with the boundary of the preset circular coverage area of ​​all edge nodes in the three-dimensional virtual scene model, filter out the intersection point closest to the starting point, and determine the circular coverage area where the intersection point is located as the coverage area of ​​the next hop edge node that the vehicle is about to enter.

[0100] S54. Extract the real-time speed and heading angle of vehicles from traffic situation characteristics, read the vehicle identification, package the vehicle identification, continuous movement trajectory curve, real-time speed and heading angle of vehicles, and current state attributes of virtual vehicle models into a data package to generate a handover data package.

[0101] In this embodiment, S6 specifically includes:

[0102] S61. Establish communication links between edge nodes, read the vehicle identification and the predicted coverage area of ​​the next-hop edge node from the handover data packet, call the preset cooperative communication protocol interface, encapsulate the handover data packet into a transmission data frame, send it to the next-hop edge node, and at the same time read the queue instruction uploaded to the cloud center, parse the target storage address in the instruction, read the regular traffic flow data into a binary stream, and send it to the cloud center through the transmission control protocol.

[0103] S62. Read the real-time computing resource status parameters of the edge computing node, and concatenate the CPU utilization, memory remaining amount and network bandwidth load into a three-dimensional resource status vector in numerical order. Input the three-dimensional resource status vector into the input layer of the improved squeezing and excitation residual network, and use a convolution kernel of preset size to perform sliding window convolution operation on the three-dimensional resource status vector along the time step to extract the preliminary resource feature map.

[0104] S63. Construct a residual main branch within the improved squeeze and excitation residual network. Input the preliminary resource feature map into the first convolutional unit. Set the first convolutional unit to contain 64 convolutional kernels. Perform dimensionality-up convolution on the preliminary resource feature map using a convolution operation with a preset stride of 1. Input the output features into the second convolutional unit. Set the second convolutional unit to contain 32 convolutional kernels. Perform dimensionality-down convolution on the features using a convolution operation with a preset stride of 1. Generate a residual feature map and add it element-wise to the preliminary resource feature map to obtain a fused feature map that integrates deep and shallow information.

[0105] S64. After the residual main branch of the improved squeezing and excitation residual network, set a bilinear pooling branch in parallel, read the fused feature map and make a copy to obtain a copy feature map, calculate the product of the transpose matrix of the fused feature map and the copy feature map, normalize each value in the product result matrix by dividing it by the total number of spatial pixels of the fused feature map, and generate a second-order interactive feature matrix that captures the second-order interactive information of the feature channels.

[0106] S65. Perform global average pooling on the second-order interaction feature matrix along the spatial width and height dimensions to compress the two-dimensional matrix into a one-dimensional channel description vector and input it into the first fully connected layer. The preset number of neurons is one-quarter of the number of channels. Perform linear dimensionality reduction on the one-dimensional channel description vector, perform ReLU activation, and input it into the second fully connected layer to restore the dimension to the original number of channels. Then, use the Sigmoid activation function to generate highly robust weight recalibration parameters.

[0107] S66. Multiply the weight recalibration parameters with the fused feature map according to the corresponding positions of each channel, recalibrate the weight of each channel of the fused feature map, read the preset resource load over-limit threshold, calculate the difference between the current CPU utilization and the resource load over-limit threshold, if the difference is greater than zero, determine that the load is over-limit, and trigger the precise evaporation operation of non-critical data.

[0108] S67. Respond to the precise evaporation operation of non-critical data, traverse the standardized edge input dataset in the local cache, read the dynamic control weight set value corresponding to each handover data packet, preset the data evaporation threshold, and filter out data packets with dynamic control weight values ​​less than the data evaporation threshold.

[0109] S68. Mark the selected data packets as non-critical data, delete the original vehicle video stream and radar point cloud data in the non-critical data packets, and extract and retain only the vehicle bounding box coordinates, vehicle type labels, vehicle real-time speed and heading angle. Encapsulate the extracted data into a simplified data stream that retains only structured feature data.

[0110] S69. Read the virtual vehicle model in the local micro digital twin, read the three-dimensional position coordinates of the virtual vehicle model according to the preset sampling interval, calculate the difference of the three-dimensional position coordinates at adjacent sampling times, count the queue length values ​​of all virtual vehicle models, combine the three-dimensional position coordinate differences with the queue length values, and output the current running status.

[0111] This implementation method achieves intelligent monitoring of edge computing resources and precise evaporation of non-critical data by introducing an improved squeezing and excitation residual network combined with a bilinear pooling mechanism. Data is diverted to the next-hop node or the cloud using a cooperative communication protocol, and preliminary resource features are extracted through convolutional operations. In particular, the constructed bilinear pooling branch generates highly robust weights by capturing second-order interaction information of feature channels, effectively enhancing the model's sensitivity to changes in resource status. When the load exceeds limits, the system accurately identifies and removes the original video stream and point cloud data based on dynamically managed weights, retaining only key structured features. This method significantly reduces storage and transmission pressure while ensuring real-time status monitoring of local microscopic digital twins, significantly improving the system's operating efficiency and data management capabilities under high-load scenarios.

[0112] The improved squeezed and excited residual network of this invention is similar to the original squeezed and excited residual network in that both retain the core architecture of the residual network, namely, extracting features through convolutional layers and using skip connections to add the input and output element by element to fuse deep and shallow information. Both also use global average pooling to compress the spatial dimension and use fully connected layers to learn channel weights.

[0113] The difference lies in that this invention breaks the limitation of the original squeeze and excitation residual network, which only captures the first-order interaction information between feature channels, and introduces a bilinear pooling mechanism. Building upon the original model's direct global pooling and channel dimensionality reduction, this invention sets up a bilinear pooling branch in parallel in step S64. By calculating and normalizing the product of the fused feature map and its transpose, a second-order interaction feature matrix capturing the second-order interaction information of the feature channels is generated. Subsequently, in step S65, a fully connected layer with a preset ratio is used to perform a linear transformation and sigmoid activation on the second-order interaction matrix, generating highly robust weight recalibration parameters.

[0114] The beneficial effect of the improvements lies in the fact that by introducing a bilinear pooling mechanism, the improved squeezed and excited residual network can accurately capture subtle second-order feature dependencies between channels, breaking the limitation of the original squeezed and excited residual network which only uses first-order statistical information for modeling, and achieving refined management of feature recalibration. This design significantly enhances the model's ability to represent complex states of edge computing resources and can generate more targeted weight parameters; when data evaporation is triggered by overload, it can more accurately identify non-critical data, significantly reducing system load while ensuring the integrity of traffic situation characteristics, and enhancing the robustness and real-time response capability of the intelligent traffic management system.

[0115] In this embodiment, S7 specifically includes:

[0116] S71. Read the running status and simplified data stream of the local microscopic digital twin, extract the lane queue length value contained in the running status, and extract the vehicle bounding box coordinates, vehicle type label, vehicle real-time speed and heading angle contained in the simplified data stream. Then, concatenate the lane queue length value and the extracted vehicle feature data according to the preset data structure order.

[0117] S72. Encapsulate the spliced ​​data into a control data packet of a preset format, and send the control data packet to the data receiving interface of the traffic control equipment at the intersection to complete the closed-loop control of traffic data on the edge side.

[0118] A smart transportation data management and control system based on edge computing includes:

[0119] Multi-dimensional traffic data synchronous acquisition and resource perception module: used to synchronously acquire multi-dimensional traffic perception data at intersections and obtain real-time computing resource status parameters of the current edge computing node;

[0120] Multi-source data standardization and fusion module: used to perform real-time background modeling and foreground segmentation on vehicle video streams in multi-dimensional traffic perception data, map radar point cloud data to a unified geospatial coordinate system, and align it with vehicle visual features in time and space to generate a standardized edge input dataset.

[0121] Dynamic weight generation module: It is used to input the standardized edge input dataset into the edge computing node, extract potential value representations by using a multi-view spatiotemporal interactive attention network, and generate a dynamic control weight set by calculating the data entropy value and time sensitivity of the current traffic scenario and combining it with real-time computing resource status parameters.

[0122] Adaptive traffic splitting module: It is used to implement adaptive traffic splitting on the standardized edge input dataset according to the dynamic control weight set. For data with high weight and containing complex road condition features, it performs real-time structured processing and extracts traffic situation features at the local edge node. For low-weight regular traffic flow data, it compresses and encapsulates the data and generates queue instructions.

[0123] Digital twin mapping and handover processing module: used to map traffic situation features to the local micro-digital twins constructed inside the edge nodes, and predict the coverage area of ​​the next-hop edge node based on the vehicle movement trajectory, and generate handover data packets;

[0124] The Deep Feature Extraction and Precise Evaporation Control Module is used to send the handover data packet to the predicted next-hop edge node, respond to queue instructions to upload regular traffic flow data to the cloud center, and use an improved squeezing and excitation residual network for feature extraction. A bilinear pooling mechanism is introduced in the residual branch to capture the second-order interaction information of the feature channel, generate weight recalibration parameters, trigger the precise evaporation operation of non-critical data when the load exceeds the limit, generate a simplified data stream, and output the real-time monitoring status of the local micro digital twin.

[0125] Control data output module: Used to read the running status and simplified data stream, splice them in the order of the preset data structure, encapsulate them into a control data packet of preset format and output them.

[0126] Example 1: To verify the feasibility of this invention in the field of intelligent transportation data management, it was applied to the "Intelligent Transportation Digital Twin Management Platform" in a national-level intelligent connected vehicle test demonstration zone in a certain city. This demonstration zone covers 45 kilometers of urban main roads and highway connectors, and is equipped with 128 roadside intelligent sensing units (RSUs), 360 high-definition AI cameras, and 210 millimeter-wave radar nodes. It receives massive amounts of multi-source heterogeneous data in real time, including vehicle trajectory data, traffic signal control status, road environment video streams, and vehicle-road cooperative interaction information, with a daily peak data processing volume of up to 8.6TB. In traditional traffic management systems, all raw high-definition video streams collected by roadside equipment are uploaded to the cloud center without filtering, leading to severe congestion of the communication network during morning and evening rush hours. This obstructs the transmission of critical low-latency traffic signal commands, easily causing delays in intersection signal control and triggering regional traffic congestion or even traffic accidents.

[0127] In practical applications of this scenario, this invention fully leverages the advantages of edge computing nodes in data source processing. Through intelligent management and control modules deployed within roadside units, it achieves refined distribution and diversion of massive traffic data. The system first establishes highly reliable communication links between roadside edge nodes, the cloud center, and neighboring nodes, classifying real-time collected data packets according to vehicle identification and business priority. For critical data related to life safety, such as emergency rescue and vehicle collision warnings, the system encapsulates it into high-priority transmission frames through a pre-defined collaborative communication protocol interface, sending them to the next-hop edge node with minimal latency via a dedicated channel or directly uploading them to the cloud, ensuring millisecond-level emergency response. For massive amounts of routine traffic flow monitoring videos and background environmental data from non-congested road sections, the system proceeds to deep feature extraction and resource assessment. The system reads the CPU utilization, remaining memory, and network bandwidth load of edge computing nodes in real time, constructing a three-dimensional resource state vector, and inputting it into an improved squeezing and incentive residual network based on a bilinear pooling mechanism.

[0128] This network extracts deep features through residual main branches and uses bilinear pooling to capture second-order interaction information between feature channels, generating weight recalibration parameters that accurately reflect data importance. When the system determines that the current edge node's load exceeds the limit, it immediately triggers a precise evaporation operation for non-critical data. The system traverses the standardized edge input dataset in the local cache, filters out data packets with low dynamic control weights, deletes the original high-definition video streams and dense point cloud data that occupy a large amount of storage space, and extracts and retains only core structured features such as vehicle bounding box coordinates, vehicle type labels, vehicle real-time speed, and heading angle. Subsequently, these simplified data streams are fused with the virtual vehicle model in the local microscopic digital twin. The system reads the three-dimensional position coordinates of the virtual vehicle model according to a preset sampling interval, calculates the difference between adjacent sampling times and counts the queue length, and finally outputs the fused running status to the cloud, ensuring that the cloud digital twin platform can still accurately reproduce the real traffic operation situation even with a very small amount of data. Table 1 below shows the comparison data between the method of this invention and the traditional method during the training period:

[0129] Table 1. Performance Comparison between Edge Computing Intelligent Traffic Management System and Traditional Centralized Management Mode

[0130]

[0131] Based on the detailed comparison data shown in Table 1, it can be seen that the intelligent traffic data management and control system based on edge computing proposed in this invention has significant performance advantages over the traditional centralized management and control mode in actual traffic scenarios.

[0132] In terms of data transmission efficiency, this invention reduces the amount of data actually transmitted to the cloud by an average of 77% through a precise non-critical data evaporation mechanism. This means that a massive amount of invalid video data is successfully intercepted at the edge, greatly freeing up communication network bandwidth. The peak bandwidth utilization of the transmission network drops significantly from around 98% in the traditional mode to below 40%, completely solving the network congestion problem.

[0133] In terms of response timeliness, the average response latency of critical commands has been drastically reduced from nearly 2,000 milliseconds (2 seconds) in the traditional mode to 10.5 milliseconds, achieving a nearly 100-fold improvement. This millisecond-level response speed is crucial for the safe guidance of autonomous vehicles and the real-time optimization of traffic signals.

[0134] Furthermore, due to the highly structured and simplified nature of the uploaded data, the average time for data retrieval in the cloud has been reduced from over 55 seconds to about 6 seconds, improving retrieval efficiency by nearly 9 times and saving nearly 78% of storage resources, significantly reducing the system's hardware investment and operation and maintenance costs.

[0135] In summary, this invention successfully achieves efficient management and control of intelligent transportation big data through intelligent feature extraction and bilinear pooling mechanism at the edge, providing an innovative technical solution for building an efficient, low-consumption, and practical modern intelligent transportation system.

[0136] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart transportation data management and control method based on edge computing, characterized in that, Includes the following steps: S1. Synchronously collect multi-dimensional traffic perception data at intersections and obtain real-time computing resource status parameters of the current edge computing node; S2. Perform real-time background modeling and foreground segmentation on vehicle video streams in multi-dimensional traffic perception data, extract pixel-level masks and visual features, map radar point cloud data to a unified geospatial coordinate system, and align it with vehicle visual features in time and space to generate a standardized edge input dataset. S3. Input the standardized edge input dataset into the edge computing node, use the multi-view spatiotemporal interactive attention network to extract potential value representations, calculate the data entropy value and time sensitivity of the current traffic scenario, and combine real-time computing resource status parameters to generate a dynamic control weight set. S4. Based on the dynamic control weight set, implement adaptive traffic splitting processing on the standardized edge input dataset. For data with high weight and containing complex road condition features, perform real-time structured processing and extract traffic situation features at the local edge node. For low-weight regular traffic flow data, compress and encapsulate it and generate queue instructions. S5. Map traffic situation features to a local micro-digital twin constructed inside the edge node, and predict the coverage area of ​​the next-hop edge node based on the vehicle movement trajectory to generate a handover data packet. S6. The handover data packet is sent to the predicted next-hop edge node. The response queue command uploads regular traffic flow data to the cloud center. An improved squeezing and excitation residual network is used for feature extraction. A bilinear pooling mechanism is introduced into the residual branch to capture second-order interaction information of feature channels, generating weight recalibration parameters. When the load exceeds limits, a precise evaporation operation of non-critical data is triggered, generating a simplified data stream. The real-time output monitors the operating status of the local microscopic digital twin. S7. Read the running status and simplified data stream, and concatenate them according to the preset data structure order. Then, encapsulate them into a control data packet of preset format and output it.

2. The intelligent transportation data management and control method based on edge computing according to claim 1, characterized in that, S1 specifically includes: synchronously collecting multi-dimensional traffic perception data through sensing devices deployed at intersections. The multi-dimensional traffic perception data includes vehicle video streams, radar point cloud data, and vehicle speed and location information uploaded by roadside units. The system monitoring interface of the edge computing node is called to read the current operating indicators and obtain the real-time computing resource status parameters of the edge computing node. The real-time computing resource status parameters include CPU utilization, remaining memory, and network bandwidth load.

3. The intelligent transportation data management and control method based on edge computing according to claim 1, characterized in that, S2 specifically includes: S21. Read the vehicle video stream from the multidimensional traffic perception data, use the Gaussian mixture model to statistically model the pixels in the continuous video frames, calculate the difference between the pixel value of the current frame and the mean of the background model, determine the pixel area with a difference greater than the preset difference threshold as the foreground area, and extract the pixel-level mask of the vehicle target. S22. Perform a convolution operation of a preset size on the image area covered by the pixel-level mask, calculate the gradient magnitude and direction values ​​pixel by pixel, and generate visual features. S23. Acquire radar point cloud data. Based on the image features and point cloud coordinates of the calibration board in the public field of view collected in advance, use the PnP algorithm to solve the spatial transformation relationship between the radar coordinate system and the camera coordinate system. Then, calculate the rotation matrix and translation vector by minimizing the reprojection error through nonlinear optimization. Map the radar point cloud data to a unified geospatial coordinate system so that the point cloud coordinates and the pixel coordinates in the vehicle video stream are established in correspondence. S24. Based on the correspondence, project the radar point cloud data onto the image plane where the visual features are located, and align the times of the two data acquisitions according to the timestamps to generate a standardized edge input dataset.

4. The intelligent transportation data management and control method based on edge computing according to claim 1, characterized in that, S3 specifically includes: S31. Input the standardized edge input dataset into the edge computing node to construct a multi-view spatiotemporal interactive attention network. Preset the first feature encoding branch and the second state encoding branch for parallel input. Input the standardized edge input dataset into the first feature encoding branch and the real-time computing resource state parameters into the second state encoding branch. Perform tensor concatenation operation on the outputs of the two branches. S32. Inside the first feature encoding branch, a one-dimensional temporal convolution kernel of a preset size is set, and a sliding window is performed on the standardized edge input dataset along the time axis. The dot product value of the feature vectors in the window is calculated and scaled by dividing by the square root of the vector dimension. It is then converted into a probability value through the Softmax function to generate a temporal dependent weight matrix and multiplied with the standardized edge input dataset to extract data content features. S33. Inside the second state coding branch, a multi-layer perceptron structure containing a preset fully connected layer is constructed. The CPU utilization, memory remaining and network bandwidth load in the real-time computing resource state parameters are multiplied by a preset weight matrix and superimposed with a bias vector. After the first linear transformation, the ReLU activation function is input for nonlinear mapping and then after the second linear transformation, the edge resource state features are extracted. S34. Using data content features as query vectors and edge resource status features as key and value vectors, calculate the product of the query vector and the transpose of the key vector to obtain the attention score matrix, and input it into the Softmax function for normalization. Multiply the result with the value vector matrix to establish a global association between data content features and edge resource status features, and generate a potential value representation. S35. Perform statistical distribution analysis on the potential value representation, count the number of samples of different data categories in the traffic scenario, divide by the total number of samples to calculate the probability of occurrence of each category, multiply the probability of occurrence by the natural logarithm function and sum over all categories to obtain the data entropy value, and at the same time read the generated timestamp, calculate the difference between the current timestamp and the generated timestamp, divide the difference by the preset time decay constant to obtain the timeliness sensitivity. S36. Construct a weight generation network containing two preset fully connected layers. Concatenate the data entropy value, time sensitivity, and real-time computing resource status parameters on the feature dimension, input them into the first fully connected layer, multiply them by preset weights and add biases, process them through the Tanh activation function and input them into the second fully connected layer, and map the output result to the zero to one interval through the Sigmoid activation function to generate a dynamically controlled weight set.

5. The intelligent transportation data management and control method based on edge computing according to claim 1, characterized in that, S4 specifically includes: S41. Read the weight value corresponding to each standardized edge input dataset in the dynamic control weight set, compare the weight value with the preset high weight diversion threshold. If the weight value is greater than or equal to the high weight diversion threshold, it is determined to be high weight data; otherwise, it is determined to be low weight data. S42. For high-weight data, read vehicle video stream images from the standardized edge input dataset, perform convolution operations using a pre-defined convolutional neural network, extract feature maps, and generate vehicle bounding box coordinates and vehicle type labels through regression calculation. S43. Simultaneously read radar point cloud data, calculate the Euclidean distance between point clouds, group points with a distance less than the preset clustering threshold into one class, calculate the position change of the cluster center point, divide by the corresponding time interval to obtain the real-time speed and heading angle of the vehicle, and splice the vehicle bounding box coordinates, vehicle type label, real-time speed and heading angle of the vehicle according to the preset data structure order to extract traffic situation features. S44. For low-weight data, read the vehicle video stream from the standardized edge input dataset, use the H.264 coding standard to perform discrete cosine transform on the image frames in the video stream, quantize the transformed frequency coefficients, discard high-frequency coefficients, and retain low-frequency coefficients. S45. Simultaneously read radar point cloud data, preset the size of the voxel grid, divide the point cloud space into multiple cubic grids, calculate the average coordinates of the point cloud in each grid, replace all the original point cloud coordinates in the grid with the average coordinates, package and encapsulate the processed video stream data and point cloud data in binary data format, write the target storage address in the header of the encapsulated data, and generate a queue instruction to upload to the cloud center.

6. The intelligent transportation data management and control method based on edge computing according to claim 1, characterized in that, S5 specifically includes: S51. Create a 3D virtual scene model containing road geometric topology information as a local micro digital twin, read traffic situation characteristics, map vehicle bounding box coordinates and vehicle type labels to the corresponding positions in the 3D virtual scene model, and generate a virtual vehicle model consistent with the real-time traffic state. S52. In the three-dimensional virtual scene model, read the position coordinates of the same vehicle at consecutive times in the order of timestamps, calculate the Euclidean distance vector between the position coordinates at adjacent times, and connect all the Euclidean distance vectors end to end to form the continuous movement trajectory curve of the vehicle. S53. Read the coordinates of the end point of the continuous movement trajectory curve as the starting point, obtain the instantaneous movement direction of the vehicle at the starting point, construct a ray of preset length, calculate the intersection point of the ray with the boundary of the preset circular coverage area of ​​all edge nodes in the three-dimensional virtual scene model, filter out the intersection point closest to the starting point, and determine the circular coverage area where the intersection point is located as the coverage area of ​​the next hop edge node that the vehicle is about to enter. S54. Extract the real-time speed and heading angle of vehicles from traffic situation characteristics, read the vehicle identification, package the vehicle identification, continuous movement trajectory curve, real-time speed and heading angle of vehicles, and current state attributes of virtual vehicle models into a data package to generate a handover data package.

7. The intelligent transportation data management and control method based on edge computing according to claim 1, characterized in that, S6 specifically includes: S61. Establish communication links between edge nodes, read the vehicle identification and the predicted coverage area of ​​the next-hop edge node from the handover data packet, call the preset cooperative communication protocol interface, encapsulate the handover data packet into a transmission data frame, send it to the next-hop edge node, and at the same time read the queue instruction uploaded to the cloud center, parse the target storage address in the instruction, read the regular traffic flow data into a binary stream, and send it to the cloud center through the transmission control protocol. S62. Read the real-time computing resource status parameters of the edge computing node, and concatenate the CPU utilization, memory remaining amount and network bandwidth load into a three-dimensional resource status vector in numerical order. Input the three-dimensional resource status vector into the input layer of the improved squeezing and excitation residual network, and use a convolution kernel of preset size to perform sliding window convolution operation on the three-dimensional resource status vector along the time step to extract the preliminary resource feature map. S63. Construct a residual main branch within the improved squeeze and excitation residual network. Input the preliminary resource feature map into the first convolutional unit. Set the first convolutional unit to contain 64 convolutional kernels. Perform dimensionality-up convolution on the preliminary resource feature map using a convolution operation with a preset stride of 1. Input the output features into the second convolutional unit. Set the second convolutional unit to contain 32 convolutional kernels. Perform dimensionality-down convolution on the features using a convolution operation with a preset stride of 1. Generate a residual feature map and add it element-wise to the preliminary resource feature map to obtain a fused feature map. S64. After the residual main branch of the improved squeezing and excitation residual network, set a bilinear pooling branch in parallel, read the fused feature map and make a copy to obtain a copy feature map, calculate the product of the transpose matrix of the fused feature map and the copy feature map, normalize each value in the product result matrix by dividing it by the total number of spatial pixels of the fused feature map, and generate a second-order interactive feature matrix. S65. Perform global average pooling on the second-order interaction feature matrix along the spatial width and height dimensions to compress the two-dimensional matrix into a one-dimensional channel description vector and input it into the first fully connected layer. The preset number of neurons is one-quarter of the number of channels. Perform linear dimensionality reduction on the one-dimensional channel description vector, perform ReLU activation, and input it into the second fully connected layer to restore the dimension to the original number of channels and generate weight recalibration parameters through the Sigmoid activation function. S66. Multiply the weight recalibration parameters with the fused feature map according to the corresponding positions of each channel, recalibrate the weight of each channel of the fused feature map, read the preset resource load over-limit threshold, calculate the difference between the current CPU utilization and the resource load over-limit threshold, if the difference is greater than zero, determine that the load is over-limit, and trigger the precise evaporation operation of non-critical data. S67. Respond to the precise evaporation operation of non-critical data, traverse the standardized edge input dataset in the local cache, read the dynamic control weight set value corresponding to each handover data packet, preset the data evaporation threshold, and filter out data packets with dynamic control weight values ​​less than the data evaporation threshold. S68. Mark the selected data packets as non-critical data, delete the original vehicle video stream and radar point cloud data in the non-critical data packets, and extract and retain only the vehicle bounding box coordinates, vehicle type labels, vehicle real-time speed and heading angle. Encapsulate the extracted data into a simplified data stream that retains only structured feature data. S69. Read the virtual vehicle model in the local micro digital twin, read the three-dimensional position coordinates of the virtual vehicle model according to the preset sampling interval, calculate the difference of the three-dimensional position coordinates at adjacent sampling times, count the queue length values ​​of all virtual vehicle models, combine the three-dimensional position coordinate differences with the queue length values, and output the current running status.

8. The intelligent transportation data management and control method based on edge computing according to claim 1, characterized in that, Specifically, S7 includes: S71. Read the running status and simplified data stream of the local microscopic digital twin, extract the lane queue length value contained in the running status, and extract the vehicle bounding box coordinates, vehicle type label, vehicle real-time speed and heading angle contained in the simplified data stream. Then, concatenate the lane queue length value and the extracted vehicle feature data according to the preset data structure order. S72. Encapsulate the spliced ​​data into a control data packet of a preset format, and send the control data packet to the data receiving interface of the traffic control equipment at the intersection to complete the closed-loop control of traffic data on the edge side.

9. A smart traffic data management and control system based on edge computing, comprising executing the smart traffic data management and control method based on edge computing as described in any one of claims 1 to 8, characterized in that, include: Multi-dimensional traffic data synchronous acquisition and resource perception module: used to synchronously acquire multi-dimensional traffic perception data at intersections and obtain real-time computing resource status parameters of the current edge computing node; Multi-source data standardization and fusion module: used to perform real-time background modeling and foreground segmentation on vehicle video streams in multi-dimensional traffic perception data, map radar point cloud data to a unified geospatial coordinate system, and align it with vehicle visual features in time and space to generate a standardized edge input dataset. Dynamic weight generation module: It is used to input the standardized edge input dataset into the edge computing node, extract potential value representations by using a multi-view spatiotemporal interactive attention network, and generate a dynamic control weight set by calculating the data entropy value and time sensitivity of the current traffic scenario and combining it with real-time computing resource status parameters. Adaptive traffic splitting module: It is used to implement adaptive traffic splitting on the standardized edge input dataset according to the dynamic control weight set. For data with high weight and containing complex road condition features, it performs real-time structured processing and extracts traffic situation features at the local edge node. For low-weight regular traffic flow data, it compresses and encapsulates the data and generates queue instructions. Digital twin mapping and handover processing module: used to map traffic situation features to the local micro-digital twins constructed inside the edge nodes, and predict the coverage area of ​​the next-hop edge node based on the vehicle movement trajectory, and generate handover data packets; The Deep Feature Extraction and Precise Evaporation Control Module is used to send the handover data packet to the predicted next-hop edge node, respond to queue instructions to upload regular traffic flow data to the cloud center, and use an improved squeezing and excitation residual network for feature extraction. A bilinear pooling mechanism is introduced in the residual branch to capture the second-order interaction information of the feature channel, generate weight recalibration parameters, trigger the precise evaporation operation of non-critical data when the load exceeds the limit, generate a simplified data stream, and output the real-time monitoring status of the local micro digital twin. Control data output module: Used to read the running status and simplified data stream, splice them in the order of the preset data structure, encapsulate them into a control data packet of preset format and output them.

Images