Urban road network traffic monitoring management method

Abstract

The invention provides an urban road network traffic monitoring management method, and relates to the technical field of urban road network traffic monitoring management. The method comprises the following steps: acquiring monitoring multivariate heterogeneous data through monitoring equipment distributed on an urban road network, and preprocessing the monitoring multivariate heterogeneous data based on calibration parameters and time-space synchronization information of the monitoring equipment to obtain target multivariate heterogeneous data; performing feature extraction on the target multivariate heterogeneous data to obtain multi-dimensional semantic feature representation; based on the multi-dimensional semantic feature representation, carrying out attribute recognition on vehicles on the urban road network through a cascaded classification regression algorithm to obtain vehicle recognition data; and carrying out parameter extraction on space-time trajectory data of the vehicle based on the vehicle identification data to obtain urban road network parameter data. According to the scheme, the accuracy of vehicle identification and urban road network parameter extraction is improved.

Description

Technical Field

[0001] The present application relates to the technical field of urban road network traffic monitoring and management, and specifically to an urban road network traffic monitoring and management method. Background Art

[0002] With the acceleration of urbanization and rapid economic development, the traffic volume of urban road networks has increased exponentially. The types of vehicles on the road have increased, including cars, buses, trucks, etc., with different driving speeds and behavior patterns. At the same time, the urban road network environment has become more and more complex. In addition to the normal urban road network flow, it is also often affected by weather changes (such as heavy rain, fog, heavy snow, strong light, etc.), road construction, temporary control and other factors. When faced with these complex situations, traditional urban road network traffic monitoring and management methods often have problems such as inaccurate monitoring data, lack of timeliness and effectiveness of management measures.

[0003] At present, commonly used urban road network monitoring technologies include induction coil detection and video monitoring. Although induction coil detection can obtain vehicle passing information to a certain extent, it is easily affected by factors such as road damage and vehicle type, and may fail in bad weather (such as waterlogging the coil). Video monitoring technology is greatly restricted by light conditions. At night, in low light, strong light reflection (such as direct sunlight), etc., the image quality decreases, affecting the accuracy of vehicle identification and urban road network parameter extraction. Summary of the invention

[0004] The present application provides a method for monitoring and managing urban road network traffic, which can improve the accuracy of vehicle identification and urban road network parameter extraction.

[0005] In a first aspect, the present application provides a method for monitoring and managing urban road network traffic, the method comprising: Acquire monitoring multivariate heterogeneous data through monitoring devices distributed on the urban road network, and pre-process the monitoring multivariate heterogeneous data based on calibration parameters and spatiotemporal synchronization information of the monitoring devices to obtain target multivariate heterogeneous data; Extracting features of the target multivariate heterogeneous data to obtain a multi-dimensional semantic feature representation; Based on the multi-dimensional semantic feature representation, the attributes of vehicles on the urban road network are identified by a cascaded classification and regression algorithm to obtain vehicle identification data; Parameters of the spatiotemporal trajectory data of the vehicle are extracted based on the vehicle identification data to obtain urban road network parameter data.

[0006] By adopting the above technical solution and deploying various types of monitoring equipment on the urban road network, multivariate heterogeneous monitoring data including images, point clouds, radars, and trajectories can be obtained. The calibration parameters and spatiotemporal synchronization information of the monitoring equipment are used to calibrate the coordinate system, timestamp, interpolate, and synchronize the data frames of the original data, which can effectively eliminate the spatiotemporal differences between data from different sources and obtain the target multivariate heterogeneous data in a unified format. This lays the foundation for subsequent data fusion and collaborative processing.

[0007] After obtaining the target multivariate heterogeneous data, this method further extracts features from it to obtain a multi-dimensional semantic feature representation. Among them, a deep convolutional neural network is used to detect vehicles and extract 2D features from image data, a 3D point cloud semantic segmentation algorithm is used to extract 3D geometric features of the point cloud, and radar speed measurement data and spatiotemporal trajectory data are combined to extract speed and motion features. By fusing these features of different modalities and different levels, vehicle characteristics can be characterized from multiple dimensions such as appearance, geometry, semantics, and behavior, forming a comprehensive and compact feature representation.

[0008] Based on multi-dimensional semantic feature representation, this method uses a cascade classification and regression algorithm to accurately identify vehicle attributes. The algorithm first performs preliminary identification of features through a cascade classifier, and then performs logical verification, confidence evaluation and screening on the identification results, which can effectively improve the accuracy and reliability of vehicle attribute identification. Compared with traditional single feature recognition, the multi-dimensional semantic features formed by fusing multi-source data can complement and verify each other, overcome the limitations of single features, and maintain high recognition accuracy even in harsh environments.

[0009] Finally, based on the vehicle identification data, this method extracts multi-level and multi-granular urban road network parameters. At the micro level, it extracts parameters such as the speed, acceleration, and headway of individual vehicles; at the meso level, it extracts parameters such as the occupancy rate of lanes and sections, the average speed, and the queue length; at the macro level, it extracts parameters such as the urban road network flow, density, and OD distribution of the entire road network. In addition, dynamic parameters such as the real-time trajectory and behavioral events of vehicles are also extracted. Through the organic combination of multi-level parameters, the spatiotemporal evolution of urban road network flow can be comprehensively and three-dimensionally portrayed, and the microscopic mechanism behind the macroscopic urban road network phenomenon can be explored.

[0010] In a second aspect of the present application, a city road network traffic monitoring and management system is provided, comprising: A data acquisition module is used to acquire monitoring multivariate heterogeneous data through monitoring devices distributed on the urban road network, and pre-process the monitoring multivariate heterogeneous data based on calibration parameters and spatiotemporal synchronization information of the monitoring devices to obtain target multivariate heterogeneous data; A feature extraction module is used to extract features of the target multi-dimensional heterogeneous data to obtain a multi-dimensional semantic feature representation; A vehicle identification module, used to identify the attributes of vehicles on the urban road network through a cascaded classification and regression algorithm based on the multi-dimensional semantic feature representation to obtain vehicle identification data; The urban road network parameter acquisition module is used to extract parameters of the vehicle's spatiotemporal trajectory data based on the vehicle identification data to obtain urban road network parameter data.

[0011] In a third aspect of the present application, a computer storage medium is provided, wherein the computer storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the above method steps.

[0012] In the fourth aspect of the present application, an electronic device is provided, including a processor, a memory, a user interface and a network interface, the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes the above method.

[0013] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. This application deploys various types of monitoring equipment on the urban road network to obtain multivariate heterogeneous monitoring data including images, point clouds, radars, trajectories, etc. The calibration parameters and spatiotemporal synchronization information of the monitoring equipment are used to calibrate the coordinate system, timestamp, interpolate and synchronize the data frames of the original data, which can effectively eliminate the spatiotemporal differences between data from different sources and obtain the target multivariate heterogeneous data in a unified format. This lays the foundation for subsequent data fusion and collaborative processing.

[0014] 2. After obtaining the target multivariate heterogeneous data, this method further extracts features from it to obtain a multi-dimensional semantic feature representation. Among them, a deep convolutional neural network is used to detect vehicles and extract 2D features from image data, a 3D point cloud semantic segmentation algorithm is used to extract 3D geometric features of the point cloud, and radar speed measurement data and spatiotemporal trajectory data are combined to extract speed and motion features. By fusing these features of different modalities and different levels, vehicle characteristics can be characterized from multiple dimensions such as appearance, geometry, semantics, and behavior to form a comprehensive and compact feature representation.

[0015] 3. This application is based on multi-dimensional semantic feature representation. This method uses a cascaded classification and regression algorithm to accurately identify vehicle attributes. The algorithm first performs a preliminary identification of features through a cascade classifier, and then performs logical verification, confidence assessment and screening on the identification results, which can effectively improve the accuracy and reliability of vehicle attribute identification. Compared with traditional single feature recognition, the multi-dimensional semantic features formed by integrating multi-source data can complement and verify each other, overcome the limitations of single features, and maintain a high recognition accuracy even in harsh environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flow chart of a method for monitoring and managing urban road network traffic provided in an embodiment of the present application; Figure 2 An architecture diagram of a city road network traffic monitoring and management system provided in an embodiment of the present application; Figure 3 A schematic diagram of the structure of an electronic device provided in this application. DETAILED DESCRIPTION

[0017] In order to enable technicians in this field to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0018] In the description of the embodiments of the present application, words such as "for example" or "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "for example" or "for example" is intended to present related concepts in a specific way.

[0019] In the description of the embodiments of the present application, the meaning of the term "multiple" refers to two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals. In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "include", "comprise", "have" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0020] In order to facilitate understanding of the method and system provided by the embodiments of the present application, before introducing the embodiments of the present application, the background of the embodiments of the present application is first introduced.

[0021] With the acceleration of urbanization and rapid economic development, the number of motor vehicles has exploded, and the pressure on urban road networks has increased. The continuous increase in urban road network traffic has led to frequent congestion problems, which not only affects travel efficiency, but also brings environmental pollution and safety hazards. At the same time, the types of vehicles on the road are becoming more and more diverse, with cars, buses, trucks and other different types of vehicles mixed on the road. Their speed characteristics and behavior patterns are significantly different, which further increases the complexity of urban road network management.

[0022] In addition to the heterogeneity of traffic flow itself, the complex and changeable urban road network environment also puts higher requirements on monitoring and management. Severe weather conditions, such as heavy rain, heavy fog, and heavy snow, will significantly affect the normal driving of vehicles and road traffic capacity. Factors such as road construction and temporary urban road network control also often lead to sudden changes in urban road network traffic and path diversion. External interference factors such as strong light exposure and insufficient light will also affect the data collection quality of monitoring equipment. Faced with such complex and changeable scenarios, traditional urban road network traffic monitoring and management methods are often stretched to the limit and difficult to adapt to actual needs. Problems such as insufficient data accuracy, low real-time performance, and lagging management measures are becoming increasingly prominent, and a smarter, more accurate, and more comprehensive solution is urgently needed.

[0023] In the current urban road network monitoring technology, induction coil detection and video surveillance are the two mainstream methods. The induction coil is buried under the road and obtains traffic information by sensing the changes in the magnetic field when the vehicle passes. Although this method can provide data on the time and number of vehicles passing, its limitations are also obvious. First, the coil is easily affected by factors such as road damage, vehicle weight and chassis height differences, resulting in reduced detection accuracy. Secondly, severe weather, especially heavy rain and water accumulation, can cause the coil to be submerged and unable to work properly. In addition, the coil can only provide vehicle passing information and cannot obtain richer data dimensions, such as vehicle model, speed, trajectory, etc.

[0024] Video surveillance technology uses cameras to collect images of urban road networks and implements vehicle detection, tracking and identification through computer vision algorithms. However, video surveillance is easily restricted by external lighting conditions. Insufficient light conditions such as nighttime and rainy days can significantly reduce image quality, resulting in missed or false detections of vehicle detection and identification. Factors such as strong direct light and light reflection can also cause image overexposure or partial information loss. In addition, the data obtained by video surveillance is limited to the two-dimensional image information of the vehicle, lacking an accurate description of the vehicle's motion state, and making it difficult to support refined urban road network parameter extraction.

[0025] After the background introduction of the above content, those skilled in the art can understand the problems existing in the prior art. The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0026] Based on the above background technology, please refer to Figure 1 , Figure 1A flow chart of a method for monitoring and managing urban road network traffic provided in an embodiment of the present application. The system can be implemented by a computer program or can be run as an independent tool application. Specifically, in an embodiment of the present application, the method can be applied on a server, but can also be applied to electronic devices such as a server. A method for monitoring and managing urban road network traffic includes the following steps: S101, acquiring monitoring multivariate heterogeneous data through monitoring devices distributed on the urban road network, and preprocessing the monitoring multivariate heterogeneous data based on calibration parameters and spatiotemporal synchronization information of the monitoring devices to obtain target multivariate heterogeneous data; Specifically, after obtaining the original monitoring multivariate heterogeneous data, this application needs to perform necessary preprocessing to solve the inconsistency problem of data from different sources in time and space scales. First, the calibration parameters of each monitoring device are used to perform coordinate system conversion and spatial alignment on the original data. Through the internal and external parameter matrices obtained by calibration, the image pixel coordinates, radar polar coordinates, etc. are converted to a unified world coordinate system to achieve spatial alignment of data of different modalities. At the same time, high-precision time synchronization devices, such as GPS timing, NTP servers, etc., are used to synchronize and calibrate the clocks of each device to ensure that the data collected by different sensors remain consistent in timestamps.

[0027] After spatial registration and time synchronization, the sampling frequency and data frame rate of different sensor data are further aligned. Due to the different working principles and hardware characteristics of various sensors, their data update frequencies may vary. In order to achieve frame-level fusion of data, it is necessary to unify different sensor data to the same time base through methods such as interpolation or resampling. For example, the data frequency of millimeter-wave radar can be increased to match the video frame rate, or the video data can be aligned with the lidar data through interpolation between frames.

[0028] After a series of preprocessing operations such as coordinate system conversion, time synchronization, and data frame alignment, the original monitoring multivariate heterogeneous data is converted into target multivariate heterogeneous data with consistent time and space and unified format. Such data not only retains the unique information of each modality, but also realizes the precise registration of data in the time and space dimensions, laying a solid foundation for subsequent data fusion and collaborative processing. The effective fusion of multi-source heterogeneous data can make up for the shortcomings of a single sensor and provide more comprehensive and accurate urban road network monitoring information. For example, video data can provide the appearance characteristics and license plate information of the vehicle, and radar data can provide the speed and trajectory information of the vehicle. The combination of the two can realize the accurate identification and tracking of the vehicle.

[0029] Based on the above embodiment, as an optional embodiment, the monitoring multivariate heterogeneous data is preprocessed based on the calibration parameters and spatiotemporal synchronization information of the monitoring device to obtain the target multivariate heterogeneous data, including: S201, performing coordinate system one processing on the monitoring multivariate heterogeneous data based on calibration parameters to obtain first monitoring multivariate heterogeneous data; Specifically, the calibration parameters describe the spatial position and posture relationship of the sensor relative to a reference coordinate system (such as the world coordinate system), including parameters such as position (translation vector) and angle (rotation matrix). For each sensor, its calibration parameters are obtained in advance through offline calibration or online self-calibration. Then, using these parameters, the monitoring data collected by the sensor is transformed from the coordinate system of the sensor itself to the reference coordinate system. Taking the video camera as an example, the image pixel coordinates can be converted to the three-dimensional space coordinates in the world coordinate system through the camera imaging model and calibration parameters; for radar, the polar coordinates or rectangular coordinates are unified to the reference coordinate system consistent with other sensors through coordinate rotation and translation transformation.

[0030] By unifying the coordinate system of all heterogeneous sensor data, the monitoring data is converted into the same "semantic space", and the data of different modalities obtain spatial consistency, laying the foundation for subsequent data fusion and joint interpretation. Unified coordinate system processing not only improves the interpretability and availability of data, but also helps to explore the intrinsic connections between different sensor data and achieve information complementarity and enhancement. For example, video data provides more intuitive and semantically rich environmental information, but is greatly affected by factors such as lighting and occlusion; while radar data is not sensitive to these factors and can provide accurate distance and speed information. After unifying them into the same coordinate system, video semantics and radar accuracy can be organically combined to achieve stable perception all day and all weather.

[0031] After the coordinate system is processed, the first monitoring multivariate heterogeneous data obtained realizes the alignment and matching of different modal data in the spatial dimension, eliminates the coordinate system differences of the original data, and lays the foundation for the integrated processing and fusion application of data. This not only improves the timeliness and continuity of the data, but also expands the breadth and depth of data mining. Based on the data of the unified coordinate system, a more comprehensive and accurate representation of urban road network scenes can be established, the complementarity and correlation of multimodal data can be explored, the performance of urban road network state perception, prediction, and decision-making can be improved, and the management of smart urban road networks can be promoted into a new stage.

[0032] S202, performing time stamp calibration on the first monitoring multivariate heterogeneous data based on the spatiotemporal synchronization information to obtain second monitoring multivariate heterogeneous data; Specifically, the spatiotemporal synchronization information describes the corresponding relationship between different sensor data in the time dimension, including parameters such as clock deviation and transmission delay. These parameters can be obtained through special synchronization equipment (such as GPS timing module), or estimated through data-driven methods, such as cross-correlation analysis and feature matching. Using the known spatiotemporal synchronization information, the first monitoring multivariate heterogeneous data are timestamped one by one, and their timestamps are converted from their respective clock systems to the same reference clock. Usually, the timestamp of a sensor (such as a GPS-synchronized camera) is selected as a reference benchmark, and the data timestamps of other sensors are offset-corrected relative to this benchmark. After correction, the data of different sensors obtain synchronization in the time dimension, and the heterogeneous data corresponding to the same moment can be accurately matched.

[0033] Through timestamp calibration, the time consistency of the first monitoring multivariate heterogeneous data is guaranteed, and the different modal data are accurately aligned on the time axis to form a unified time series. This is not only conducive to data management and retrieval, and improves the real-time response capability of the system, but also provides the necessary premise for data fusion. On the basis of time synchronization, the multi-time observation data of the same target from different sensor perspectives can accurately correspond, the spatiotemporal trajectory of the target can be seamlessly spliced, and data redundancy and conflict can be eliminated. On the contrary, if there is no timestamp calibration, the uncertainty of asynchronous data will cause deviations in the fusion results, affecting the performance of perception and prediction. For example, in the target tracking task, if the detection results of the video and radar cannot be aligned in time, it is difficult to reliably estimate the state and trajectory of the target.

[0034] The second monitoring multivariate heterogeneous data obtained after calibration takes into account the consistency of spatial coordinates and timestamps, and different modal data are synchronized and matched in both time and space. On the one hand, this time-space consistency reduces the uncertainty of the data, improves the interpretability and reliability of the data, and provides a better foundation for data mining; on the other hand, time-space synchronization is also the key to heterogeneous data fusion. Accurate time-space alignment helps to deeply explore the complementarity of multimodal information, realize information enhancement at the semantic level, and build a more comprehensive and accurate understanding of the urban road network environment. For example, in the detection of urban road network events, time-space synchronized video and radar data can verify each other. Video provides semantic interpretation of events, and radar provides accurate speed and trajectory information. The combination of the two can significantly improve the accuracy of detection and recognition.

[0035] S203, performing time interpolation processing on the second monitoring multivariate heterogeneous data based on the time-space synchronization information to obtain third monitoring multivariate heterogeneous data; Specifically, time interpolation processing refers to estimating the data value at an unknown time based on known data points, so that the discrete data sequence becomes continuous and smooth. Common interpolation methods include linear interpolation, spline interpolation, Gaussian process regression, etc. For the second monitoring multivariate heterogeneous data, a unified time reference and interpolation frequency are first determined based on the spatiotemporal synchronization information, and then the data sequence of each sensor is interpolated independently. Taking linear interpolation as an example, for two consecutive sampling times t1 and t2 of a certain sensor, and the corresponding data values ​​x1 and x2, the data value x(t)=x1+(x2-x1)*(t-t1) / (t2-t1) at any time t between t1 and t2 can be estimated by linear interpolation. Through linear interpolation, data can be evenly filled between the original sampling points, making the data sequence continuous and smooth in time.

[0036] By performing time interpolation on the second monitoring multivariate heterogeneous data, the time resolution of the data is improved, and the discrete sampling points on the time axis are continuous, forming a data series with consistent time and uniform intervals, namely the third monitoring multivariate heterogeneous data. This continuous and smooth time series is conducive to mining the time correlation and evolution law of the data and improving the accuracy of data analysis and modeling. On the basis of time continuity, heterogeneous data can be accurately aligned on the time axis, the observation data of different sensors can be seamlessly spliced, and data missing and inconsistency can be compensated. On the contrary, if there is no time interpolation processing, discrete and uneven data will bring difficulties to time series analysis and state estimation, resulting in oscillation and instability in the results. For example, in the task of urban road network flow prediction, if the input urban road network flow data is discrete and uneven in time, it is difficult for the prediction model to accurately capture the dynamic change law of urban road network flow, and the prediction accuracy will be affected.

[0037] The third monitoring multivariate heterogeneous data achieves continuity and smoothness in the time dimension, while maintaining the spatial coordinate unification and timestamp synchronization of the first two steps. The spatiotemporal consistency and continuity of the data provide a basis for the fusion analysis of multi-source heterogeneous data, making it possible to mine continuous patterns from discrete data.

[0038] S204, unifying the data frames of the third monitoring multivariate heterogeneous data to obtain target multivariate heterogeneous data.

[0039] Specifically, data frame unification refers to converting heterogeneous data of different sources and types into a unified data representation form, so that they are consistent and comparable at the semantic level. Here, "data frame" is used as a unified data representation unit. Each data frame contains a timestamp and a set of synchronous observation data from heterogeneous sensors. For video data, each frame of the image can be encapsulated into a data frame together with its metadata; for radar data, the point cloud or target list at each moment can be encapsulated into a data frame; other types of sensor data can also be processed similarly. Through the encapsulation of data frames, heterogeneous data of different modalities obtain a unified logical representation, they correspond one to one on the time axis, and the semantic information they carry is also unified.

[0040] After the data frame is unified, the third monitoring multi-heterogeneous data is converted into structured target multi-heterogeneous data. The heterogeneous data from different sources are stored and organized in a unified format, which provides the basis for interoperability and association analysis. This unified data representation is conducive to data management and quality control, and improves the efficiency of data access and query. On the basis of semantic consistency, the relevance of heterogeneous data at the content level is highlighted, and the multi-dimensional information carried by different modal data can be organically integrated to form a more comprehensive and three-dimensional understanding of the urban road network scene. On the contrary, if there is no data frame consistency, it is difficult to establish semantic connections between fragmented heterogeneous data, and data fusion can only stay at a shallow level of splicing, making it difficult to exert the synergistic gain of multi-source data. For example, in multi-target tracking tasks, if the target information of video and radar cannot be semantically aligned, it is difficult to achieve cross-modal target matching and trajectory association, and redundant observations of multiple sensors are difficult to use to improve tracking accuracy.

[0041] S102, extracting features of target multi-dimensional heterogeneous data to obtain multi-dimensional semantic feature representation; Specifically, this embodiment uses a deep convolutional neural network to extract features from image data. First, the image data is input into a pre-trained CNN model, such as ResNet, Inception, etc., and multi-scale and multi-level features of the image are extracted through operations such as convolution and pooling. These features depict the hierarchical abstraction of the image content, from low-level texture and edge information to high-level semantics and contextual information. Then, based on the CNN model, the supervision information of vehicle detection and attribute recognition is integrated, and the feature representation is optimized through an end-to-end learning method to focus more on the semantic description of the vehicle target. The extracted image features not only contain geometric information such as the location and scale of the vehicle, but also high-level semantics such as vehicle model, license plate, and color.

[0042] For the three-dimensional point cloud data collected by the lidar, this embodiment uses a point cloud semantic segmentation algorithm for feature learning. Unlike traditional point cloud segmentation methods, semantic segmentation not only divides the point cloud into different clusters, but also assigns a semantic label to each cluster, such as vehicles, pedestrians, buildings, etc. Through end-to-end deep learning models such as PointNet++ and SqueezeSegV3, the original point cloud is gradually abstracted into a high-dimensional feature vector while retaining the geometric structure and semantic attributes of the point cloud. In the feature learning process, vehicle prior knowledge is introduced, and the model's ability to distinguish vehicle point clouds is improved through strategies such as sample weighting and difficult example mining. In the end, the point cloud features not only depict the three-dimensional contour and posture of the vehicle, but also integrate the category and component information of the vehicle.

[0043] For millimeter-wave radar data, this embodiment extracts the micro-motion trajectory characteristics of the vehicle. Radar data records the vehicle's motion state in the form of discrete points, with a high time resolution, which is suitable for describing the microscopic dynamic characteristics of the vehicle. First, the radar points are tracked and smoothed by the Kalman filter algorithm to obtain the vehicle's motion trajectory fragments. Then, a series of kinematic parameters such as the vehicle's speed, acceleration, steering angle, etc. are extracted from the trajectory fragments, and the vehicle's micro-motion trajectory feature vector is constructed in combination with environmental factors such as the road and urban road network flow state. This feature reflects the vehicle's motion change pattern in a short period of time, such as changing lanes, overtaking, braking, etc., which is of great significance for understanding the vehicle's behavioral intentions.

[0044] After extracting the semantic features of multi-source heterogeneous data, this embodiment organizes them into a unified multi-dimensional feature representation. On the one hand, the semantic features of different modalities are aligned in the spatiotemporal dimensions to ensure the spatiotemporal consistency of the features; on the other hand, the semantic features of different modalities complement each other in the vector space, reflecting the diversity of semantic representation. Image features characterize the visual appearance attributes of the vehicle, point cloud features characterize the geometric structure characteristics of the vehicle, and radar features characterize the microscopic motion characteristics of the vehicle, which describe the static-dynamic characteristics of the vehicle from different aspects. Combining these features into a joint representation can comprehensively characterize the semantic connotation of the vehicle from multiple dimensions, providing richer and more accurate prior knowledge for subsequent fusion recognition.

[0045] Based on the above embodiment, as an optional embodiment, the target multi-dimensional heterogeneous data includes: image data, point cloud data, radar data and spatiotemporal trajectory data; extracting features of the target multi-dimensional heterogeneous data to obtain a multi-dimensional semantic feature representation includes: S301, extracting features from the image data based on a preset deep convolutional neural network to obtain 2D feature data; Specifically, a deep convolutional neural network is a hierarchical feature learning model that can automatically learn hierarchical feature representations in image data by stacking multiple convolutional layers and pooling layers. For image data of urban road network scenes, we preset a CNN model structure suitable for semantic understanding of urban road networks, and pre-trained it on a large-scale urban road network image dataset to enable it to have the ability to recognize urban road network elements such as vehicles, pedestrians, and roads. Given a frame of urban road network monitoring image, it is first input into the preset CNN model, and after a series of convolution, activation, pooling and other operations, the local and global features of the image are extracted layer by layer. The low-level features of CNN correspond to basic visual elements such as edges and textures in the image, while the high-level features correspond to more abstract semantic concepts such as vehicle type and pedestrian posture. By intercepting the feature map at a certain intermediate layer of CNN, a compact feature vector, that is, the 2D feature representation of the image data, can be obtained.

[0046] Mapping image data to 2D feature space can filter out redundancy and noise in the original image, highlight the characteristic patterns related to the semantics of the urban road network, and significantly reduce the data dimension. This semantically rich and moderately dimensional 2D feature representation is conducive to subsequent data analysis and pattern recognition tasks. Based on the similarity measurement of 2D features, image-level urban road network event retrieval and anomaly detection can be achieved; based on the classification model of 2D features, the semantic content of urban road network scenes can be automatically annotated and understood. On the contrary, if the original image pixels are directly used as features, they are easily disturbed by factors such as lighting and viewing angle, and it is difficult to depict the content at the semantic level. For example, in the task of urban road network accident detection, it is difficult to accurately distinguish between accident vehicles and normal vehicles using pixel-level features, while using 2D semantic features makes it easier to achieve robust accident recognition.

[0047] Based on the above embodiment, as an optional embodiment, feature extraction is performed on the image data based on a preset deep convolutional neural network to obtain 2D feature data, including: Based on a preset deep convolutional neural network, the vehicle target is detected from the image data, and the 2D feature of the vehicle target is extracted to obtain 2D feature data.

[0048] Specifically, an optimized deep convolutional neural network model is used to detect targets on image data, and potential vehicle targets are quickly located from the image. Generally, a target detection network pre-trained with a large number of vehicle data sets is selected, such as a faster regional convolutional neural network, a real-time object detection model based on single multi-frame detection, etc., which fully utilizes the technical advantages of convolutional neural networks in automatic feature extraction and multi-scale feature fusion, generates dense candidate regions from feature maps of multiple scales, and uses a region proposal network to quickly screen out the region most likely to contain a vehicle from the candidate region, and obtain a series of bounding boxes of vehicle targets. In order to further improve the detection accuracy, a feature enhancement module based on the attention mechanism is introduced. By adaptively weighting the convolutional feature map, the significant features of the vehicle target are highlighted, and background interference is suppressed, making the subsequent region screening more "focused". At the same time, a difficult sample mining strategy is adopted to specifically enhance the learning of suspected vehicle targets and further compress false detections and missed detections in the detection results. For example, for scene images with densely parked vehicles, the network model is prone to missed detection. At this time, the training proportion of this type of image can be increased to enhance the network's adaptability to dense vehicles. By allowing the target detection network to continuously optimize itself in massive scenarios, its detection capabilities are constantly iteratively evolved, ultimately resulting in a robust and efficient vehicle detection model.

[0049] After accurately locating the vehicle target, the preset feature extraction network is further used to extract 2D features of the vehicle area to obtain compact and discriminative 2D feature data. Generally, convolutional neural networks pre-trained on large datasets such as imagenet are used, such as residual networks, densely connected convolutional networks, etc., and the vehicle bounding box area is input into the feature extraction network. The vehicle image is mapped to the semantic space layer by layer using the convolutional neural network to form hierarchical and structured vehicle features. In the feature extraction process, in order to fully explore the discriminative information of the vehicle's 2D appearance, on the one hand, a multi-scale feature fusion mechanism is introduced to capture the appearance details of the vehicle at different granularities through the cascade fusion of cross-layer features; on the other hand, regional average pooling, saliency weighting and other strategies are introduced for the vehicle area to highlight the key features of the vehicle and suppress redundant information such as the vehicle's internal texture. Finally, the multi-layer feature map is mapped to a fixed-length feature vector through global average pooling, L2 normalization and other operations, that is, the 2D feature data of the vehicle target. By integrating the "detection-description" process in an end-to-end manner and relying on the powerful representation capabilities of deep learning, this 2D feature extraction solution can quickly and accurately form a detailed characterization of vehicle targets. The 2D feature data can not only comprehensively reflect the visual attributes of the vehicle's appearance, but also has strong discrimination and generalization capabilities.

[0050] S302, extracting features from the point cloud data based on a 3D point cloud semantic segmentation algorithm to obtain 3D feature data; Specifically, given a frame of point cloud data of an urban road network scene, the point cloud semantic segmentation algorithm is first used to annotate it point by point. This algorithm usually adopts an end-to-end deep neural network structure, takes the original point cloud as input, and predicts the semantic category of each point through point-by-point feature learning and context fusion. The point cloud semantic segmentation network is trained on a large-scale 3D urban road network scene dataset to learn to establish a mapping relationship between the geometric features of the point cloud and the semantic concepts of the urban road network. After semantic segmentation, each point in the point cloud is assigned a clear category label, forming a semantically rich point cloud representation. On this basis, by performing feature aggregation on the point cloud of each semantic category, a set of 3D semantic features can be obtained, which characterizes the distribution of geometric features of the point cloud in different semantic categories, that is, 3D feature data.

[0051] Through 3D point cloud semantic segmentation, originally homogeneous discrete points are given heterogeneous category attributes, and the semantic information of point cloud data is highlighted. Semantically rich point cloud representation provides an important basis for scene understanding and urban road network element analysis. Based on 3D semantic features, the segmentation, positioning and interaction analysis of objects of different categories in the scene can be realized; by aggregating point clouds of objects of the same category, the three-dimensional size, posture and motion state of the object can be estimated. On the contrary, if there is a lack of semantic segmentation of point clouds, it is difficult to accurately distinguish different elements in complex urban road network scenes by relying solely on geometric features, and it is difficult to support semantic-level urban road network event analysis. For example, in the analysis of urban road network congestion, if vehicles and buildings cannot be distinguished from point clouds, it is difficult to estimate traffic density and congestion status through point cloud density.

[0052] 3D point cloud semantic segmentation realizes the transformation from raw geometric point cloud to structured semantic scene, so that the urban road network knowledge hidden in massive point cloud data can be mined and a 3D feature representation under semantic guidance can be formed. This semantic-oriented point cloud feature extraction is the cornerstone of refined perception and 3D reconstruction of smart city road network scenes. On the one hand, semantic point cloud provides data support for 3D visualization and immersive analysis of urban road network scenes, helping to build a 3D smart city road network cockpit; on the other hand, the semantic features of point cloud also provide a link for the semantic association of multimodal data, so that heterogeneous urban road network data such as point cloud, image, and trajectory can be matched and registered at the semantic level. For example, by associating the 3D point cloud features of a vehicle with the 2D image features, the pixel-level positioning and segmentation of the vehicle in the image can be achieved; by fusing the 3D motion features of the vehicle with the trajectory data, more accurate vehicle motion state estimation and abnormal behavior detection can be achieved.

[0053] Based on the above embodiment, as an optional embodiment, feature extraction is performed on point cloud data based on a 3D point cloud semantic segmentation algorithm to obtain 3D feature data, including: S501, processing the point cloud data based on a 3D point cloud semantic segmentation algorithm to obtain sub-cloud data with semantic labels; Specifically, an optimized 3D point cloud semantic segmentation algorithm is used to process the point cloud data, and the point cloud subset belonging to the vehicle target is quickly identified from the point cloud scene. Generally, a semantic segmentation network pre-trained with a large-scale vehicle point cloud dataset is selected, such as PointNet++, RandLA-Net, etc., which makes full use of the technical advantages of deep learning in automatic extraction of point cloud features and multi-scale context aggregation, and realizes pixel-level annotation of point clouds through point-by-point feature learning. In the segmentation process, in order to further strengthen the discriminative features of the vehicle area, a local feature aggregation module is introduced. By combining the local geometric structure of the vehicle point cloud, the feature expression of key vehicle components (such as wheels, lights, etc.) is strengthened, making the semantic segmentation more refined and accurate. At the same time, in view of the irregularity and disorder of the point cloud, strategies such as data augmentation and feature normalization are adopted to enhance the robustness of the network model to environmental changes and improve the adaptability of the vehicle point cloud in terms of scale, angle, etc. For example, a random rotation transformation is applied to the vehicle point cloud to make the segmentation model more invariant to the vehicle orientation. Through continuous data-driven optimization, the generalization and robustness of the semantic segmentation network are continuously improved, ultimately achieving accurate semantic segmentation of vehicle point clouds.

[0054] On the basis of semantic annotation of point cloud, 3D features are further extracted from the point cloud subset with "vehicle" label to obtain compact and discriminative 3D feature data. Generally, semantic segmentation post-processing modules such as Vehicle-Net and PointNet are used to input the semantically labeled vehicle point cloud subset into the feature extraction network. The multi-level and sparse representation of the vehicle structure is realized through point-by-point feature transformation, and the 3D geometric properties of the vehicle are characterized by a set of ordered feature vectors. In order to fully explore the discriminative information of the vehicle's 3D shape, the vehicle prior knowledge is incorporated into the feature extraction process. By introducing feature constraints based on the vehicle's 3D skeleton, the significant features of the vehicle's joints are highlighted, and the influence of the vehicle's local deformation is suppressed, so that the 3D features are more focused on the overall structure of the vehicle. At the same time, strategies such as posture normalization and learning rate annealing are adopted to enhance the posture invariance and convergence stability of the vehicle's 3D features. Finally, the multi-level vehicle point cloud features are mapped to a fixed-dimensional feature space through operations such as maximum pooling and L2 normalization to form a 3D feature vector of the vehicle target. Thanks to the semantically driven vehicle point cloud segmentation, the 3D feature extraction scheme can efficiently and accurately form a fine characterization of vehicle targets, which not only simplifies the complexity of the point cloud scene, but also retains the geometric details of the key areas of the vehicle, making the 3D features achieve a good balance in terms of discriminability and compactness.

[0055] S502, performing expansion extraction from the sub-cloud data to obtain 3D geometric feature data; Specifically, based on the semantic segmentation of vehicle point clouds, an extended 3D feature extraction method is used to learn and extract the 3D geometric features of vehicles from the subset of vehicle point clouds. Generally, typical point cloud feature learning networks such as PointNet++ and DGCNN are selected to map the vehicle point cloud to the semantic feature space through point-by-point feature transformation and hierarchical aggregation. In order to enhance the network model's ability to express the geometric shape of the vehicle, a priori constraints based on the vehicle's spatial structure are introduced in the feature extraction process, and a clearer semantic orientation is given to the vehicle point cloud by fusing the vehicle's 3D skeleton information. For example, the position priors of key parts of the vehicle (such as wheels, windows, etc.) are used as additional feature inputs to guide the network to focus on learning the discriminative parts of the vehicle and highlight the overall structural characteristics of the vehicle's geometric appearance. At the same time, in view of the sparsity and incompleteness of the vehicle point cloud, strategies such as data augmentation and feature completion are adopted to simulate partial occlusion and near-far changes of vehicle point clouds in real scenes by adding random missing and local perturbations, so as to enhance the robustness of point cloud features from the data level. In addition, when aggregating multi-level features, the vehicle spatial scale prior is incorporated, and the aggregation strength of different semantic levels is adaptively adjusted to highlight the representation of key areas of the vehicle, suppress interference from secondary areas such as the rear of the vehicle, and further enhance the discriminability of the vehicle's overall geometric appearance. Finally, the vehicle point cloud features at different semantic levels are mapped into a fixed-length 3D geometric feature vector through operations such as pooling and splicing. Thanks to the synergistic enhancement of vehicle spatial prior and data-driven feature learning, the 3D geometric feature can not only accurately characterize the spatial structural properties of the vehicle, but also has stronger semantic directivity and robustness, and can maintain good feature discriminability even when the vehicle point cloud is incomplete or sparse.

[0056] S503 , integrating the 3D geometric feature data according to the semantic labels of the sub-cloud data to obtain 3D feature data.

[0057] Specifically, based on the extraction of the 3D geometric features of the vehicle, the semantic segmentation labels of the point cloud are used to realize the semantic-driven integration of the vehicle's 3D features through adaptive weighted fusion. First, the significance weights of each region are preset according to the importance of different semantic regions to the overall features of the vehicle. Generally, the body, wheels and other regions with strong discrimination of vehicle models are given larger weights, and non-critical regions such as background and windows are given smaller weights. Then, the 3D geometric features of different semantic regions are multiplied element by element with the corresponding weights to obtain the semantically weighted regional feature vectors. On this basis, the attention mechanism is further introduced to highlight the personalized appearance characteristics of the vehicle by adaptively adjusting the weight coefficients of each region. For example, for different models, there are often obvious differences in exterior decorations such as headlights and air intake grilles. The attention network can automatically identify these difference regions and increase their weights, which can further enhance the discrimination of the vehicle's 3D features. Finally, the weighted regional feature vectors are integrated by element-wise summation to obtain semantically enhanced vehicle 3D feature data. Compared with simple multi-view feature stitching, this semantically driven 3D feature integration scheme makes full use of the semantic prior of point cloud. Through adaptive weighting of regional saliency, it highlights the discriminative details of key areas of the vehicle while lightweighting the vehicle feature representation, which significantly improves the semantic directionality and feature compactness of 3D features.

[0058] S303, extracting features from the radar data to obtain velocity feature data; Specifically, feature extraction of radar data mainly includes three steps: data preprocessing, target detection and speed estimation. In the data preprocessing stage, the original radar data is first filtered and denoised to remove random noise and clutter interference and improve the signal-to-noise ratio. Common radar data filtering methods include median filtering and Kalman filtering, which can effectively suppress noise and smooth the radar velocity curve. The preprocessed radar data enters the target detection stage, and the moving targets in the radar field of view are automatically detected through algorithms such as constant false alarm rate detection (CFAR), and the state parameters such as the distance, speed and angle of the target are estimated. For the detected targets, the Doppler effect or pulse accumulation methods are used to further extract the target's velocity characteristics. By smoothing and fusing the target velocity in time and space, stable and reliable velocity feature data is finally obtained.

[0059] Speed ​​is a key indicator of the operation status of urban road networks and the core semantic connotation of radar data. Through feature extraction of radar data, speed information is highlighted from noisy echo signals to form a structured spatiotemporal feature representation of speed. Based on speed feature data, functions such as vehicle and pedestrian speed detection and abnormal speed identification can be realized, providing data support for urban road network management and control; through statistical analysis of speed distribution of roads or intersections, the macroscopic operation characteristics of urban road network flow can be characterized, and applications such as urban road network flow prediction and congestion diagnosis can be realized. On the contrary, if the original radar data is used directly, it is easily affected by noise and clutter, and it is difficult to accurately reflect the real state of the urban road network, which poses a challenge to the data quality and the credibility of the analysis results. For example, in speeding violation detection, the use of noise-contaminated radar speed data is prone to misjudgment, while the use of feature-extracted speed data can significantly improve the accuracy of violation discrimination.

[0060] S304, extracting features from the spatiotemporal trajectory data to obtain motion feature data; Specifically, a hierarchical processing paradigm is adopted for the feature extraction of spatiotemporal trajectory data, and the physical, geometric and semantic features of the trajectory are extracted from the bottom to the top. In the physical feature extraction stage, the physical quantities such as the position, velocity and acceleration of the trajectory points are calculated to characterize the microscopic motion state of the individual in the spatiotemporal domain. Common physical features include instantaneous velocity, average velocity, velocity variance, maximum acceleration, etc., which can be estimated from the trajectory coordinate sequence by numerical methods such as finite difference. On this basis, the geometric features of the trajectory are extracted to measure the macroscopic morphological characteristics of the trajectory in the road network space. Typical trajectory geometric features include trajectory length, distance straightness, average turning angle, complexity, etc., which comprehensively reflect the geometric shape, curvature and spatial distribution range of the trajectory. Finally, domain knowledge is introduced to give semantic interpretation to the trajectory and extract high-level semantic features. For example, the trajectory is mapped to the road network topology to extract navigation semantics such as the road sections and turning relationships passed; points of interest (POI) in the trajectory are identified, and behavioral semantics such as travel purpose and activity mode of the trajectory are associated. The physical, geometric and semantic features of the trajectory are integrated to form a comprehensive and refined multi-dimensional feature description of the motion behavior, namely the motion feature data.

[0061] Through the multi-level feature extraction of spatiotemporal trajectory data, the inherent laws of individual movement are highlighted from discrete coordinate points and a structured spatiotemporal feature representation is formed. The rich trajectory movement features provide a new perspective for understanding individual urban road network behavior and analyzing group travel patterns. Based on physical features, trajectory anomaly detection and similarity retrieval can be realized; based on geometric features, trajectory shape classification and clustering can be performed; based on semantic features, high-level behavioral semantics such as travel chains and activity patterns can be inferred. On the contrary, if the analysis is directly based on the original spatiotemporal coordinate sequence, it is difficult to accurately characterize the inherent laws of movement and is easily disturbed by problems such as non-uniform sampling of trajectories and missing data. For example, in the calculation of trajectory similarity, if the Euclidean metric is used to compare the original coordinate points, trajectories that are similar in space but not synchronized in time sampling may be judged as dissimilar, while the measurement method based on physical and geometric features can more accurately judge the spatiotemporal similarity of trajectories.

[0062] S305 , fusion of 2D feature data, 3D feature data, speed feature data and motion feature data to obtain a multi-dimensional semantic feature representation.

[0063] Specifically, the feature fusion of multimodal urban road network data adopts the representation learning paradigm based on graph network. First, for each type of feature data, a homogeneous feature graph is constructed. Taking the image feature graph as an example, each image region is regarded as a node of the graph, the regional features are used as the attributes of the node, and the spatial adjacency relationship between regions is used as the edge of the graph. Similarly, point cloud feature graphs, speed feature graphs, and motion feature graphs can also be modeled as graph structures in which nodes carry corresponding semantic attributes. Then, based on the homogeneous feature graph, a heterogeneous feature graph that fuses nodes and edges of different modalities is further constructed. We use shared nodes to align semantically equivalent entities on different modal graphs, such as aligning the regional nodes of the same vehicle in the image and point cloud into the same node; then, heterogeneous semantic edges between different modalities are constructed based on physical laws and prior knowledge, such as the "speed" relationship between the vehicle node and its speed node, and the "trajectory" relationship between the vehicle node and its motion trajectory.

[0064] After constructing a heterogeneous graph that integrates multimodal features, a graph neural network is used to perform end-to-end representation learning on the heterogeneous graph. The graph neural network can learn the embedded feature representation of the nodes in the graph by performing layer-by-layer recursive feature transformation on the graph structure data, which not only integrates the semantic attributes of the nodes themselves, but also covers the correlation dependencies between nodes. In the heterogeneous graph scenario, different types of nodes and edges carry semantic knowledge of different modalities. By recursively transferring and integrating information gathered on different types of edges between nodes, the graph neural network can learn cross-modal, context-related node representations. For example, the embedded features of the vehicle node integrate both the visual semantics of the vehicle in the image and point cloud modalities, and the behavioral semantics of the vehicle in the speed and trajectory modalities, ultimately forming a multi-dimensional vehicle feature description. These node features that integrate the semantics of different modalities are further aggregated in the global scope of the graph to form a multi-dimensional semantic feature representation of the entire urban road network scene.

[0065] S103, based on the multi-dimensional semantic feature representation, attribute recognition is performed on vehicles on the urban road network by using a cascaded classification regression algorithm to obtain vehicle recognition data; Specifically, the cascaded classification-regression algorithm includes two main stages: the classification stage and the regression stage. In the classification stage, a multi-level classifier is used to divide the attribute space of the vehicle step by step. First, according to the overall characteristics of the vehicle, such as size and appearance, the vehicle is roughly divided into several categories, such as small cars, large cars, and trucks. Then, within each category, subcategories are further divided according to finer-grained features, such as lights and grilles, such as sedans, SUVs, and trucks. Through step-by-step classification, the original high-dimensional attribute space is divided into several low-dimensional subspaces, reducing the difficulty of subsequent fine recognition. At the same time, in each layer of classifiers, the attention mechanism and multi-task learning strategy are introduced to utilize the correlation between vehicle attributes, such as the model distribution of certain brands and the relationship between color and model, to improve the contextual consistency and robustness of classification decisions.

[0066] After completing the classification stage, the vehicle samples are divided into a fine-grained attribute subspace, and then enter the regression stage to accurately estimate their specific attribute values. For discrete attributes, such as vehicle model and brand, the softmax regression method is used to predict the probability distribution of each attribute value, and the attribute value with the highest confidence is taken as the final recognition result. For continuous attributes, such as vehicle speed and vehicle distance, linear regression or deep regression network is used to directly predict the numerical value of the attribute. In the regression process, the multi-dimensional semantic features of the vehicle are fully utilized, and the features of different modalities are fused to improve the accuracy of attribute estimation. For example, the three-dimensional size of the vehicle is jointly estimated by image features and point cloud features, and the speed of the vehicle is jointly estimated by image features and radar features. Through multimodal feature fusion, the shortcomings of a single feature can be complemented and the robustness of recognition can be improved. Based on the above embodiment, as an optional embodiment, based on the multi-dimensional semantic feature representation, the attributes of vehicles on the urban road network are identified by a cascaded classification regression algorithm to obtain vehicle identification data, including: S401, selecting a recognition feature subset related to a preset recognition task requirement from the multi-dimensional semantic feature representation; Specifically, for each type of vehicle attribute recognition task, we select an independent feature subspace from the multi-dimensional semantic feature representation. Taking vehicle brand recognition as an example, brand attributes mainly rely on the appearance features of the vehicle body, so a feature subset based on image features and integrated with the local point cloud features of the vehicle body is selected for brand recognition; while vehicle size attributes require an overall characterization of the vehicle shape contour, so a feature subset based on vehicle point cloud features and integrated with vehicle image contour features is selected for vehicle length and width estimation; for vehicle speed attributes, speed features and trajectory features are obviously the most relevant, so this type of feature is directly selected and a small amount of image semantics is integrated for vehicle speed recognition. The semantic feature space is selectively cropped for recognition tasks, and a customized compact feature representation is constructed for each type of task.

[0067] After obtaining the targeted feature subset, the cascade classification and regression algorithm is further used to identify the vehicle attributes. Taking vehicle type recognition as an example, the classification algorithm is first used based on the vehicle appearance features to determine the rough category of the vehicle, such as distinguishing between cars, trucks, buses, etc.; then, for each type of vehicle, the regression algorithm is used to further subdivide the specific vehicle models, such as sedans and SUVs in cars, and vans and engineering vehicles in trucks, etc. Through hierarchical classification and regression, the refined recognition of vehicle models is achieved. Similarly, license plate recognition adopts a cascade paradigm of first positioning and then recognition, and the vehicle brand and body color adopt a cascade paradigm of first classification and then regression. The organic combination of coarse-grained classification and fine-grained regression can significantly improve the overall recognition performance of vehicle attributes.

[0068] S402, inputting the recognition feature subset into a preset cascade classifier to obtain cascade recognition data; Specifically, for each type of vehicle attribute, a multi-layer cascade classifier structure is constructed. Taking vehicle type recognition as an example, the first layer uses a multi-category classifier to determine the rough category of the vehicle, such as cars, trucks, buses, etc.; the second layer deploys a dedicated fine-grained classifier for each rough category to further determine the specific vehicle type, such as the car classifier subdivided into sedans, sports utility vehicles, etc., and the truck classifier subdivided into vans, engineering vehicles, etc.; the third layer uses a regressor to estimate the specific parameters of the vehicle based on the subdivided vehicle type, such as the model and displacement of sedans, and the loading capacity and cargo category of trucks, etc. The entire recognition process is like an inverted pyramid-shaped multi-level decision, refining the vehicle attributes layer by layer from top to bottom until a comprehensive and fine-grained vehicle semantic description is obtained.

[0069] In order to improve the effect of cascade classification and recognition, vehicle prior knowledge is introduced as a constraint in the training of each layer of classifiers. For example, in vehicle model recognition, similar models such as sedans and SUVs are more likely to be confused, so they should be the same parent node in the vehicle model classification tree and share more feature semantics; while trucks and buses have large differences in appearance and should be placed in different branches of the classification tree. In addition, the feature selection of classifiers of different granularity is also integrated with priors. For example, the vehicle brand mainly relies on appearance features, so image features are selected, and vehicle attributes need to be combined with speed trajectory features. By integrating vehicle structured prior knowledge with data-driven classifier learning, vehicle semantic reasoning from coarse to fine and from qualitative to quantitative is formed.

[0070] The cropped feature subset is input into the matching cascade classifier, and after layer-by-layer reasoning, cascade recognition data of vehicle attributes is generated step by step. The top layer outputs the overall category label of the vehicle, such as cars, trucks, etc.; the middle layer outputs the subdivision labels within the category, such as sedans and sports utility vehicles among cars, vans and engineering vehicles among trucks, etc.; the bottom layer outputs the fine parameters of the vehicle, such as length, width, height, brand model, and cargo loaded. Each layer of recognition data is a further refinement of the semantic representation of the vehicle, and the overall formation of a hierarchical characterization of vehicle attributes from the surface to the inside, from coarse to fine. For example, the single-layer vehicle model recognition outputs rough categories such as cars and buses, and after cascade classification, it outputs subdivided vehicle models such as sedans, medium-sized buses, and double-decker buses, and then combines regression to output the specific model and parameters of each vehicle.

[0071] S403, performing a logic check on the cascade recognition data to obtain first cascade recognition data, and performing a confidence evaluation on the first cascade recognition data to obtain recognition confidence of the first cascade recognition data; Specifically, the cascade recognition data output by the cascade classifier is converted into a symbolic attribute knowledge graph. Taking the recognition of vehicle type and license plate number as an example, the recognition results of vans for cars and training cars for medium-sized buses are obviously contrary to common sense, so it is necessary to detect and correct such erroneous results. With the help of the existing symbolic knowledge in the vehicle attribute library, such as "cars cannot be vans" and "training cars belong to cars", the abnormal attribute combinations in the recognition knowledge graph are checked based on rules, and the recognition results with low confidence and large contradictions are eliminated. The missing attributes are supplemented by combining attribute associations to obtain the first cascade recognition data with logical self-consistency. For example, when a vehicle is identified as a car, its license plate number should meet the small car license plate specifications. If the recognized license plate number does not meet the specifications, the recognition result is corrected using logical rules such as "cars cannot have large license plates". After obtaining the first cascade recognition data, it is further evaluated for confidence to obtain the credibility of each recognition data. On the one hand, the performance differences of classifiers at different levels will lead to uneven confidence in recognition results, and the confidence needs to be quantified. On the other hand, the difficulty of distinguishing different attributes of vehicles varies, and the requirements for recognition confidence are also different. It is necessary to adaptively adjust the confidence threshold from the perspective of the task. For example, the difficulty of identifying the brand model of a vehicle is much higher than that of the body color, so the confidence threshold for brand recognition should be higher than that for color; for example, the cost of identifying a truck license plate number is higher than that for a car, so the confidence threshold for truck license plate recognition should be higher than that for a car. By setting adaptive thresholds in the recognition branches of each attribute and calculating the confidence of each attribute using methods such as Bayesian calibration, the hierarchical recognition confidence of the first-level cascade recognition data is finally obtained, providing a quantitative indicator for semantic discrimination. When the recognition confidence of an attribute is lower than the preset threshold, the recognition of the attribute can be judged as a failure, and subsequent optimization iterations are initiated.

[0072] S404, screening the first cascade recognition data according to the recognition confidence to obtain second cascade recognition data; Specifically, the screening intensity of attribute recognition is set according to the importance of vehicle attributes. For example, vehicle brand and license plate number are rigid requirements, so the screening threshold of their recognition confidence should be set high to ensure the recognition accuracy of key attributes; while the screening threshold of attributes such as body color, year and model can be appropriately relaxed to balance the efficiency and cost of recognition. At the same time, the attribute screening threshold needs to be adaptively adjusted in combination with the task type, and the screening threshold curve matching the task is dynamically generated by weighing the requirements of different tasks for recognition accuracy and real-time performance. For example, for offline tasks such as vehicle retrieval and vehicle statistics, recognition accuracy is the core demand, so the screening intensity can be relatively strict; while real-time tasks such as vehicle monitoring and urban road network induction have higher requirements for timeliness, and the screening intensity can be appropriately relaxed. By presetting the screening threshold curves of different tasks in the attribute semantic library and using the threshold adaptive algorithm to optimize the screening curves online, a closed-loop drive from task to data is finally achieved, so that vehicle attribute screening can be done "within mind". After determining the screening thresholds for each attribute recognition, classic screening methods such as Top-K are used to filter the first cascade recognition data, remove the recognition results with confidence levels below the threshold, and obtain the second cascade recognition data with high credibility and strong correlation. For example, when the confidence level of license plate number recognition is lower than 0.9, the recognition result will be screened out; when the top-3 results with the highest confidence levels in the candidate list of vehicle brands are all lower than 0.8, brand recognition is judged to have failed.

[0073] S405, performing structured processing on the second cascade identification data to obtain vehicle identification data.

[0074] Specifically, a unified structured template is defined with reference to the basic vehicle attribute library, and the attribute names, data types, etc. of the second-level cascade recognition data are standardized. Taking vehicle brands as an example, since there are various ways to express vehicle brands in different data sources and different recognition tasks, such as "BMW", "BMW", "Bayerische Motoren Werke AG", etc., it is necessary to refer to the vehicle brand name standard, merge these attribute names with synonyms, and uniformly map them into standardized brand names, such as "BMW". At the same time, there may be non-vehicle attribute data in the second-level cascade recognition data, such as driver information, environmental information, etc., which also need to be eliminated through attribute filtering. The standardized vehicle attributes are encapsulated using general structured languages ​​such as XML (Extensible Markup Language) to form a hierarchical attribute knowledge representation, such as <vehicle> <brand> BMW< / brand> <model> X5< / model> <color> black< / color> < / vehicle>On this basis, the vehicle attribute ontology library is further used to semantically associate structured attributes. By defining the logical relationship between attributes, such as "BMW is a German brand" and "X5 is an SUV", an attribute knowledge graph is constructed to achieve the transition from "information island" to "association network". For example, by associating the vehicle brand with the place of origin and the vehicle model with the level, semantic tuples such as brand-origin and model-level are formed, which constitute the basic unit of the vehicle attribute knowledge network. By using intelligent methods such as ontology reasoning machines, the deep relationship between attributes is continuously explored, such as "BMW X5 is usually black or white", and the explicit attributes of the vehicle are combined with the implicit preferences to form a three-dimensional, multi-dimensional attribute association network, that is, the final vehicle identification structured data.

[0075] S104, extracting parameters of the spatiotemporal trajectory data of the vehicle based on the vehicle identification data to obtain urban road network parameter data.

[0076] Specifically, based on the vehicle identification data, the position of each vehicle is first mapped to a specific road and lane in combination with the road map information to form a spatiotemporal trajectory sequence of the vehicle. The spatiotemporal trajectory contains the vehicle's position, speed, acceleration and other state information, as well as the corresponding timestamp and road number. After obtaining the vehicle trajectory, this embodiment aggregates and analyzes the trajectory at different spatiotemporal granularities from both macro and micro levels to extract the corresponding urban road network parameters.

[0077] At the macro level, taking roads or road sections as the basic units, all vehicle trajectories within a period of time are aggregated and counted to obtain parameters that reflect the overall urban road network status. For example, by counting the number of vehicles passing through a certain section of road per unit time, the traffic volume of the section is obtained; by calculating the average speed of passing vehicles, the average driving speed of the section is obtained; by recording the number and duration of low-speed or stopped vehicles, the delay and queue length of the section are obtained. In addition to these conventional parameters, this embodiment also introduces some novel macro parameters, such as road capacity utilization, lane balance, etc., to reflect the efficiency of road resource utilization and the rationality of traffic flow distribution. Through the extraction of macro parameters, the overall operating status of the road network can be quantitatively evaluated, providing a basis for urban road network planning, induction, control and other decision-making.

[0078] At the micro level, the vehicle is used as the basic unit to finely characterize the trajectory of each vehicle and extract parameters that reflect the interaction between vehicles. For example, by analyzing the position and speed differences of the front and rear vehicles, the headway and vehicle spacing are obtained; by counting the frequency and amplitude of the lateral displacement of the vehicle per unit time, the lane change frequency and lane change intensity are obtained; by extracting the curvature change and angular velocity of the vehicle trajectory, the turning radius and turning rate are obtained. In addition, typical driving behavior modes such as following and free flow can be identified based on the relative position and speed relationship of multiple vehicle trajectories. Through the extraction of micro parameters, the driving characteristics and interaction laws of vehicles on the road can be quantitatively described, providing basic data for studying driving behavior, predicting urban road network conflicts, and optimizing traffic organization.

[0079] On the basis of macro and micro parameter extraction, this embodiment further explores the correlation characteristics between parameters and constructs a multi-granularity parameter fusion model. On the one hand, in the time dimension, the dynamic change trend of parameters is analyzed, and time series characteristics that reflect the fluctuation law of urban road networks are extracted, such as morning and evening peak period characteristics, holiday characteristics, etc.; on the other hand, in the spatial dimension, the correlation between parameters between adjacent sections and lanes is analyzed, and collaborative patterns that reflect the spatiotemporal evolution of urban road network flows are extracted, such as intersection traffic efficiency, trunk flow distribution, etc. Through the multi-granularity fusion of time and space, not only can the dynamic change law of urban road network parameters be characterized, but also the intrinsic connection between different parameters can be revealed, providing a new perspective for a deep understanding of the mechanism of urban road networks.

[0080] After multi-level and multi-granular urban road network parameter extraction, vehicle identification data is converted into a series of structured parameter indicators, covering macro urban road network flow characteristics and micro vehicle behavior characteristics, forming a comprehensive description of the operation status of the road urban road network. These parameter data can not only be used to monitor the status of the urban road network in real time and detect abnormal events in a timely manner, but can also be used as historical data accumulation to mine urban road network laws, establish prediction models, optimize signal timing, etc., providing comprehensive data support for intelligent urban road network management and control.

[0081] Based on the above embodiment, as an optional embodiment, the spatiotemporal trajectory data of the vehicle is subjected to parameter extraction based on the vehicle identification data to obtain urban road network parameter data, including: S601, extracting micro-level urban road network parameters based on vehicle identification data to obtain micro-level urban road network parameters; Specifically, after obtaining the vehicle identification data, the vehicle's spatiotemporal trajectory is preprocessed using methods such as sliding windows and Kalman filtering to smooth the noise and abnormal points in the trajectory data and improve the data quality. Then, taking a single vehicle as the object, various parameters reflecting its microscopic driving state are extracted. The first is the speed parameter. By tracking the changes in the position of the vehicle in the previous and next two frames within a fixed time window, the vehicle's instantaneous speed, average driving speed, maximum speed, etc. are estimated. The second is the acceleration parameter. By performing the first-order difference on the speed sequence, the vehicle's instantaneous acceleration, average acceleration, etc. are obtained. The third is the headway. By tracking the joint changes in the position and speed dimensions of the front and rear vehicles, the headway is estimated by dividing the distance between points by the speed difference. The fourth is the lane change behavior. By analyzing the relative position changes between the vehicle trajectory and the lane line, the occurrence time and duration of operations such as lane change and overtaking are identified. In the extraction process, the noise interference such as positioning error and vehicle occlusion is fully considered, and the average smoothing and threshold filtering are used to improve the robustness of parameter estimation. At the same time, according to the physical meaning of the parameters, a reasonable threshold interval is set to eliminate invalid values ​​with obvious abnormalities. After vehicle-by-vehicle analysis, a parameter set reflecting the microscopic individual characteristics of urban road network flow is finally formed. Thanks to the fine characterization of vehicle trajectory data, the extracted microscopic parameters have higher spatiotemporal resolution and sampling frequency than traditional floating vehicle data, and can accurately describe the instantaneous driving state of vehicles, providing a data basis for restoring the fine structure of urban road network flow.

[0082] S602, extracting mesoscopic urban road network parameters based on vehicle identification data to obtain mesoscopic urban road network parameters; Specifically, after obtaining the vehicle identification data, the geographical location information of the trajectory point is first matched with the electronic map of the road network, and the vehicle trajectory is mapped to specific road, lane and other road section units. On the basis of road section matching, a suitable statistical time and space range is set to extract the mesoscopic parameters of the urban road network flow on each road section unit. The first is the traffic density. By counting the number of vehicles in a unit section length, the traffic density parameters of each lane and each direction are obtained to characterize the spatial distribution characteristics of the urban road network flow at the road section scale. The second is the headway distribution. The queue state of vehicles in the road section is restored by the vehicle trajectory data, and the numerical distribution law of the headway is statistically analyzed, which is generally described by statistical quantities such as mean, variance, and quantile. The third is the queue length. For sections such as signal intersections and toll stations, the start and stop status of the vehicle is determined by the trajectory data, and the queue length of the fleet, the number of overflow lanes and other parameters are extracted. The fourth is the travel time. By tracking the entry and exit time of vehicles crossing the road section, the statistical parameters such as the mean and variance of the travel time of different sections are estimated. During the extraction process, we fully consider abnormal situations such as temporary parking of vehicles and trajectory noise, remove invalid trajectory segments, and improve the accuracy of parameter estimation. At the same time, cluster analysis is used to identify the multi-state characteristics of urban road network flow, and statistical parameters under different urban road network flow states such as free flow and synchronous flow are extracted in a targeted manner. After extracting parameters from each road section, a mesoscopic parameter set that reflects the macroscopic characteristics of the road urban road network state is finally formed. Compared with microscopic individual parameters, mesoscopic parameters reflect the group characteristics of urban road network flow through cross-vehicle aggregation, which can be directly used to analyze the traffic capacity of the road network and evaluate the service level of urban road network facilities.

[0083] S603, extracting macro-level urban road network parameters based on vehicle identification data to obtain macro-level urban road network parameters; Specifically, after obtaining massive vehicle identification data, we first determine the appropriate time and space scope according to the analysis target, such as morning and evening peak hours, urban road network divisions, etc. On this basis, we summarize and count all vehicle trajectory data that fall within the analysis scope, and extract various macro parameters that reflect the operating status of the regional urban road network. The first is the total travel volume. By statistically analyzing the total number of vehicles in the time period, we can characterize the overall travel demand intensity within the region. The second is the travel distance distribution. By counting the distance between the starting and ending points of the vehicle trajectory, we can estimate parameters such as the average travel distance and travel distance distribution. The third is the peak hour coefficient. By comparing the difference in urban road network volume during peak hours and off-peak hours, we can quantitatively describe the time-varying characteristics of regional urban road network demand. The fourth is the speed distribution. By aggregated analysis of the speed data of all trajectory points in the region, we can obtain the average driving speed and speed distribution parameters of the regional road network. The fifth is the OD matrix. The starting and ending points of the vehicle are restored through the trajectory data, and integrated into the urban road network cell unit, finally forming a dynamic OD matrix of the regional scope to characterize the time and space distribution pattern of the macro urban road network flow. In the parameter extraction process, a big data analysis architecture is adopted, and distributed storage, parallel computing and other technologies are used to improve data processing efficiency. At the same time, with the help of cloud computing, edge computing and other technical means, real-time processing and online analysis of massive vehicle trajectory data are realized. After multi-level and multi-period parameter extraction, a parameter set reflecting the macro characteristics of regional urban road network operation is finally formed. Compared with traditional sampling surveys, this parameter extraction method has the advantages of large sample size, full temporal and spatial coverage, and fine analysis granularity. It can provide more comprehensive and accurate data support for regional urban road network operation evaluation, urban road network planning, policy simulation, etc.

[0084] S604, extracting dynamic urban road network parameters based on vehicle identification data to obtain dynamic urban road network parameters; Specifically, after obtaining vehicle identification data, the trajectory sequence is divided by a fixed-time sliding window to extract the change status of urban road network parameters in each sliding window. The first is the speed change rate. By tracking the change trend of vehicle speed over time in the sliding window, the mean and variance of the speed change rate of all vehicles are estimated to characterize the dynamic change level of the overall speed of the urban road network flow. The second is the acceleration change rate. By performing trend analysis on the vehicle acceleration sequence in the sliding window, the acceleration change rate distribution parameters are obtained to reflect the dynamic stability characteristics of the urban road network flow. The third is the headway time change rate. By tracking the change trend of the headway time between the front and rear vehicles in the sliding window, the headway time change rate distribution is estimated to measure the dynamic safety level of the urban road network flow. The fourth is the vehicle density change rate. By analyzing the fluctuation of vehicle density parameters in multiple consecutive time windows, the dynamic change characteristics of vehicle density are obtained. In the process of dynamic parameter extraction, data stream computing, incremental learning and other technologies are fully utilized to realize real-time processing and parameter updating of massive vehicle trajectory data. At the same time, in response to the sudden changes in the urban road network flow, such as construction control, severe weather and other scenarios, an adaptive time window and parameter threshold strategy is used to improve the adaptability and robustness of dynamic parameter extraction. After extracting dynamic parameters in multiple time periods and multiple scenarios, a set of parameters that reflect the dynamic operation characteristics of the urban road network system is finally formed. Compared with traditional static parameters, dynamic parameters reflect the inherent mechanism of the time-varying evolution of urban road network flow. Through the time series analysis of parameters, the evolution law of urban road network events can be explored, and the trend changes of urban road networks can be predicted, providing the necessary basic data support for dynamic urban road network control and travel information guidance.

[0085] S605, taking the micro-level urban road network parameters, the meso-level urban road network parameters, the macro-level urban road network parameters and the dynamic urban road network parameters as urban road network parameter data.

[0086] Specifically, firstly, a unified data storage structure is designed for urban road network parameters at different levels, such as micro, meso, macro, and dynamic, including metadata information such as parameter type, spatiotemporal granularity, and statistical range. Then, the data of each parameter is organized according to a unified structural specification to form a standardized urban road network parameter data set. On this basis, a multi-level index system of urban road network parameters is established to realize the upper and lower association and interactive retrieval of macro, meso, and micro parameters. At the same time, the spatiotemporal topological relationship of urban road network parameters is constructed to realize the joint query of static attribute parameters and dynamically changing parameters. In addition, technical means such as large-scale parallel processing and distributed storage are used to improve the management and calculation efficiency of massive urban road network parameter data. After the comprehensive integration of multi-source urban road network parameters, a set of multi-level, multi-dimensional, spatiotemporal urban road network parameter data resources is finally formed. Compared with the parameters of a single level, this parameter data reflects the complex characteristics of the urban road network system, which is a combination of macro and micro, static and dynamic. Through the correlation analysis between parameters, the laws of urban road networks can be deeply explored, the multi-scale emergence mechanism of urban road network phenomena can be revealed, and a more comprehensive and three-dimensional understanding of the urban road network system can be formed. At the same time, based on the urban road network parameter data, a set of application models for urban road network monitoring, evaluation, prediction and simulation can be constructed to support multi-scenario business needs of urban road networks.

[0087] See also Figure 2 , Figure 2 An architecture diagram of a city road network traffic monitoring and management system provided in an embodiment of the present application, the city road network traffic monitoring and management system may include: The data acquisition module 1 is used to acquire monitoring multivariate heterogeneous data through monitoring devices distributed on the urban road network, and pre-process the monitoring multivariate heterogeneous data based on the calibration parameters and spatiotemporal synchronization information of the monitoring devices to obtain target multivariate heterogeneous data; Feature extraction module 2, used for extracting features of the target multi-dimensional heterogeneous data to obtain multi-dimensional semantic feature representation; The vehicle identification module 3 is used to identify the attributes of vehicles on the urban road network through a cascaded classification and regression algorithm based on the multi-dimensional semantic feature representation to obtain vehicle identification data; The urban road network parameter acquisition module 4 is used to extract parameters of the space-time trajectory data of the vehicle based on the vehicle identification data to obtain urban road network parameter data.

[0088] It should be noted that: when the system provided in the above embodiment realizes its functions, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiment belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0089] Please refer to Figure 3 The application also discloses an electronic device. Figure 3 The electronic device 300 may include: at least one processor 301, at least one network interface 304, a user interface 303, a memory 305, and at least one communication bus 302 or end-to-end wireless communication.

[0090] The communication bus 302 is used to realize the connection and communication between these components.

[0091] The user interface 303 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0092] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0093] Among them, the processor 301 may include one or more processing cores. The processor 301 uses various interfaces and lines to connect various parts in the entire server, and executes various functions of the server and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 305, and calling data stored in the memory 305. Optionally, the processor 301 can be implemented in at least one hardware form of digital signal processing (Digital Signal Processing, DSP), field programmable gate array (Field~Programmable Gate Array, FPGA), and programmable logic array (Programmable Logic Array, PLA). The processor 301 can integrate one or a combination of a central processing unit (Central Processing Unit, CPU), a graphics processing unit (Graphics Processing Unit, GPU) and a modem. Among them, the CPU mainly processes the operating system, user interface and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 301, and it can be implemented separately through a chip.

[0094] Among them, the memory 305 may include a random access memory (Random Access Memory, RAM) and may also include a read-only memory (Read~Only Memory). Optionally, the memory 305 includes a non-transitory computer-readable medium. The memory 305 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 305 may optionally be at least one storage system located away from the aforementioned processor 301. Refer to Figure 3 , the memory 305 as a computer storage medium may include an operating system, a network communication module, a user interface module and an application program of a method for monitoring and managing urban road network traffic.

[0095] exist Figure 3In the electronic device 300 shown, the user interface 303 is mainly used to provide an input interface for the user and obtain the data input by the user; and the processor 301 can be used to call the application program storing the urban road network traffic monitoring management method in the memory 305, and when executed by one or more processors 301, the electronic device 300 executes one or more methods in the above-mentioned embodiments. It should be noted that for the aforementioned method embodiments, for the sake of simple description, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited to the described action sequence, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required for the present application. In the above embodiments, the description of each embodiment has its own emphasis, and the part that is not described in detail in a certain embodiment can refer to the relevant description of other embodiments.

[0096] In the several implementations provided in this application, it should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are only schematic, such as the division of modules, which is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interfaces, indirect coupling or communication connection of systems or modules, which can be electrical or other forms.

[0097] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0098] The present application also provides a computer storage medium that can store multiple instructions, which are suitable for being loaded and executed by a processor as described above. Figure 1 The urban road network traffic monitoring and management method of the embodiment shown in the figure can be found in the specific execution process. Figure 1 The specific description of the illustrated embodiment will not be repeated here.

[0099] In addition, each functional module in each embodiment of the present application can be integrated into a processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The above integrated modules can be implemented in the form of hardware or software functional modules.

[0100] If the integrated module is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a memory and includes several instructions for a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned memory includes: various media that can store program codes, such as USB flash drives, mobile hard drives, magnetic disks or optical disks.

[0101] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure.

[0102] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A method for monitoring and managing urban road network traffic, characterized in that: The method comprises: Acquire monitoring multivariate heterogeneous data through monitoring devices distributed on the urban road network, and pre-process the monitoring multivariate heterogeneous data based on calibration parameters and spatiotemporal synchronization information of the monitoring devices to obtain target multivariate heterogeneous data; Extracting features of the target multivariate heterogeneous data to obtain a multi-dimensional semantic feature representation; Based on the multi-dimensional semantic feature representation, the attributes of vehicles on the urban road network are identified by a cascaded classification and regression algorithm to obtain vehicle identification data; Parameters of the spatiotemporal trajectory data of the vehicle are extracted based on the vehicle identification data to obtain urban road network parameter data.

2. The method according to claim 1, characterized in that The preprocessing of the monitoring multivariate heterogeneous data based on the calibration parameters and spatiotemporal synchronization information of the monitoring device to obtain target multivariate heterogeneous data includes: Performing coordinate system one processing on the monitoring multivariate heterogeneous data based on the calibration parameters to obtain first monitoring multivariate heterogeneous data; Performing time stamp calibration on the first monitoring multivariate heterogeneous data based on the time-space synchronization information to obtain second monitoring multivariate heterogeneous data; Performing time interpolation processing on the second monitoring multivariate heterogeneous data based on the time-space synchronization information to obtain third monitoring multivariate heterogeneous data; The third monitoring multivariate heterogeneous data is unified into data frames to obtain target multivariate heterogeneous data.

3. The method according to claim 1, characterized in that The target multivariate heterogeneous data includes: image data, point cloud data, radar data and spatiotemporal trajectory data; the feature extraction of the target multivariate heterogeneous data to obtain a multi-dimensional semantic feature representation includes: Extracting features from the image data based on a preset deep convolutional neural network to obtain 2D feature data; Extracting features from the point cloud data based on a 3D point cloud semantic segmentation algorithm to obtain 3D feature data; Extracting features from the radar data to obtain velocity feature data; Extracting features from the spatiotemporal trajectory data to obtain motion feature data; The 2D feature data, the 3D feature data, the speed feature data and the motion feature data are subjected to feature fusion to obtain the multi-dimensional semantic feature representation.

4. The method according to claim 1, characterized in that The method of performing attribute recognition on vehicles on the urban road network based on the multi-dimensional semantic feature representation through a cascaded classification regression algorithm to obtain vehicle recognition data includes: Selecting a recognition feature subset related to a preset recognition task requirement from the multi-dimensional semantic feature representation; Inputting the identification feature subset into a preset cascade classifier to obtain cascade identification data; Performing a logic check on the cascade identification data to obtain first cascade identification data, and performing a confidence evaluation on the first cascade identification data to obtain identification confidence of the first cascade identification data; Screening the first cascade recognition data according to the recognition confidence to obtain second cascade recognition data; The second cascade identification data is structured to obtain the vehicle identification data.

5. The method according to claim 1, characterized in that The step of extracting parameters from the spatiotemporal trajectory data of the vehicle based on the vehicle identification data to obtain urban road network parameter data includes: Extracting micro-level urban road network parameters based on the vehicle identification data to obtain micro-level urban road network parameters; Extracting mesoscopic urban road network parameters based on the vehicle identification data to obtain mesoscopic urban road network parameters; Extracting macro-level urban road network parameters based on the vehicle identification data to obtain macro-level urban road network parameters; Extracting dynamic urban road network parameters based on the vehicle identification data to obtain dynamic urban road network parameters; The micro-level urban road network parameters, the meso-level urban road network parameters, the macro-level urban road network parameters and the dynamic urban road network parameters are used as the urban road network parameter data.

6. The method according to claim 3, characterized in that The step of extracting features from the point cloud data based on a 3D point cloud semantic segmentation algorithm to obtain 3D feature data includes: Processing the point cloud data based on the 3D point cloud semantic segmentation algorithm to obtain sub-cloud data with semantic labels; Performing extended extraction from the sub-cloud data to obtain 3D geometric feature data; The 3D geometric feature data are integrated according to the semantic labels of the sub-cloud data to obtain the 3D feature data.

7. The method according to claim 3, characterized in that The feature extraction of the image data based on a preset deep convolutional neural network to obtain 2D feature data includes: Based on the preset deep convolutional neural network, a vehicle target is detected from the image data, and 2D features of the vehicle target are extracted to obtain the 2D feature data.

8. An urban road network traffic monitoring and management system, characterized in that: The system comprises: A data acquisition module is used to acquire monitoring multivariate heterogeneous data through monitoring devices distributed on the urban road network, and pre-process the monitoring multivariate heterogeneous data based on calibration parameters and spatiotemporal synchronization information of the monitoring devices to obtain target multivariate heterogeneous data; A feature extraction module is used to extract features of the target multi-dimensional heterogeneous data to obtain a multi-dimensional semantic feature representation; A vehicle identification module, used to identify the attributes of vehicles on the urban road network through a cascaded classification and regression algorithm based on the multi-dimensional semantic feature representation to obtain vehicle identification data; The urban road network parameter acquisition module is used to extract parameters of the vehicle's spatiotemporal trajectory data based on the vehicle identification data to obtain urban road network parameter data.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: It includes a processor, a memory and a transceiver, the memory is used to store instructions, the transceiver is used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes the method according to any one of claims 1 to 7.

Images