A traffic video anomaly management and control method and system based on a three-dimensional real scene map

Abstract

The application discloses a traffic video abnormality management and control method and system based on a three-dimensional real scene map, relates to the technical field of big data analysis, and comprises the following steps: establishing three-dimensional real scene modeling of a target road; based on traffic monitoring video streams of the target road and the three-dimensional real scene modeling of the target road, generating spatial mapping of the traffic monitoring video streams and the three-dimensional real scene modeling of the target road, and marking abnormal event information of dynamic objects in the target road; performing real-time three-dimensional rendering according to the abnormal event information of the dynamic objects in the target road, analyzing abnormal risk indexes of the dynamic objects in the target road; based on the abnormal risk indexes of the dynamic objects in the target road, performing visual labeling of the three-dimensional real scene modeling of the target road, determining an abnormal event type of the dynamic objects, and generating a traffic abnormality management and control scheme of the target road. The application has the advantages of improving traffic management efficiency, reducing the probability of accidents, and optimizing traffic flow management and emergency response.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, specifically to a method and system for managing traffic video anomalies based on a 3D real-scene map. Background Technology

[0002] Traffic video anomaly control using 3D real-scene maps refers to the real-time monitoring and analysis of traffic flow, road conditions, and vehicle behavior by combining 3D map technology and video surveillance technology. This allows for the identification and early warning of abnormal situations in traffic, and the implementation of corresponding management measures to ensure traffic safety and efficiency.

[0003] Existing traffic video anomaly control mainly relies on IoT data to analyze the traffic status of the target area. However, it is prone to false alarms when dealing with large traffic flow scenarios, and can only be passively controlled based on the traffic flow status, and cannot make advance predictions and interventions based on the traffic behavior of individual targets. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides a method and system for traffic video anomaly control based on a 3D real-scene map. This technical solution solves the problems of existing traffic video anomaly control methods, which mainly rely on IoT data to analyze the traffic status of the target area. However, these methods are prone to false alarms when dealing with large traffic flow scenarios and can only be passively controlled based on the traffic flow status, without the ability to predict and intervene in advance based on the traffic behavior of individual targets.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A traffic video anomaly control method based on a 3D real-scene map includes:

[0007] S1. Based on heterogeneous traffic sensors, acquire traffic monitoring video streams and 3D modeling data of the target road, and establish a 3D real-scene model of the target road; the 3D modeling data includes: point cloud data, satellite imagery, and street view data;

[0008] S2. Based on the traffic monitoring video stream and the 3D real-scene model of the target road, generate a spatial mapping between the traffic monitoring video stream and the 3D real-scene model of the target road, and mark the abnormal event information of dynamic objects in the target road.

[0009] S3. Perform real-time 3D rendering based on the abnormal event information of dynamic objects in the target road, generate a traffic state model of the 3D real scene of the dynamic target road, and analyze the abnormal risk indicators of dynamic objects in the target road.

[0010] S4. Based on the abnormal risk indicators of dynamic objects in the target road, perform 3D real-scene modeling and visualization annotation of the target road, determine the abnormal event types of dynamic objects, and generate a traffic anomaly control plan for the target road.

[0011] Preferably, step S1 specifically includes:

[0012] Based on the multi-source data of traffic monitoring 3D modeling data of the target road, the ICP iterative nearest point algorithm is used to register the point cloud data to eliminate the bias, and satellite imagery and street view data are used as the texture source for 3D modeling to obtain the initial target road 3D real scene modeling data package.

[0013] Data preprocessing is performed on the initial 3D reality modeling data package of the target road;

[0014] Using the GDAL data conversion open-source library, coordinate unification is performed on the traffic monitoring video stream of the target road and the initial 3D real scene modeling data package of the target road to obtain multi-source fused 3D real scene modeling data of the target road;

[0015] Based on multi-source fusion of target road 3D real scene modeling data, the Poisson surface reconstruction algorithm is used to solve the gradient field of the implicit function, which is consistent with the direction of the normal vector field of the input point cloud data, to generate the target road 3D real scene model.

[0016] Specifically, the ICP iterative nearest point algorithm is as follows:

[0017] In the formula, Let be the transformation matrix. To transform the matrix The point obtained by transforming the i-th source point. For the target point cloud There are 1 point, where m is the total amount of point cloud data.

[0018] The Poisson surface reconstruction algorithm is as follows:

[0019]

[0020] In the formula, For gradient operators, Implicit function At point gradient at, For point The normal vector at that location.

[0021] Preferably, step S2 specifically includes:

[0022] Based on the traffic monitoring video stream of the target road, according to the 3D real scene model of the target road, several sets of 2D and 3D feature image data point pairs are randomly selected from the traffic monitoring video stream of the target road and substituted into the cv2.solvePnP function to calculate and solve the camera extrinsic parameters, world coordinates and intrinsic parameters of the traffic monitoring video stream of the target road, and perform spatial mapping to obtain the spatial location information of the traffic monitoring video stream of the target road.

[0023] The traffic monitoring video stream of the target road is divided according to the frame rate of the video stream to obtain the traffic monitoring image frame sequence of the target road.

[0024] Preferably, step S2 further includes:

[0025] The spatial location information of the traffic monitoring video stream of the target road is used as the image attribute in the corresponding traffic monitoring image frame sequence of the target road. The YOLOv5 algorithm is used to collect the motion trajectory of dynamic objects in the traffic monitoring image frame sequence of the target road and extract the trajectory feature data of each dynamic object. The trajectory feature data includes: speed, acceleration, and direction change.

[0026] By using DeepSORT multi-target tracking, the trajectories of dynamic objects in adjacent frames of the trajectory feature data of each dynamic object are associated, and the position coordinates of dynamic objects in the complete data stream are marked to generate a sequence of motion object trajectory frames in the traffic monitoring image of the target road.

[0027] Preferably, step S2 further includes:

[0028] Based on the Lstm long short-term memory network, an abnormal target behavior trajectory analysis model is established. The target road traffic monitoring image moving object trajectory frame sequence is used as input. According to several trajectory step lengths of the target road traffic monitoring image moving object trajectory history, the future moving object trajectory is predicted, and abnormal target behavior label data in the target road is generated.

[0029] Based on the intrinsic and extrinsic parameters of traffic monitoring sensors on the target road, the two-dimensional spatial coordinates of abnormal target behavior label data in the target road are mapped to the three-dimensional real-scene model of the target road to obtain abnormal event information of dynamic objects in the target road.

[0030] Specifically, the target behavior trajectory anomaly analysis model is as follows:

[0031]

[0032] In the formula, This refers to the abnormal event information of the j-th dynamic object in the target road at the t-th unit of time. The first in the target road The hidden state of a dynamic object at time t. The first in the target road Trajectory feature data of a dynamic object at the t-th unit of time. The first in the target road Weight matrix of a dynamic object The first in the target road Bias terms for a dynamic object It is the sigmoid activation function.

[0033] Prior to this, step S3 specifically includes:

[0034] An observation window is established according to a unit time, and the abnormal event information of dynamic objects in the target road is used as the observation attribute to extract the abnormal behavior feature data of dynamic objects in the target road.

[0035] Based on the abnormal behavior characteristic data of dynamic objects in the target road, an abnormal behavior risk quantification function is established to calculate the abnormal risk index of dynamic objects in the initial target road.

[0036] Specifically, the abnormal behavior risk quantification function is as follows:

[0037]

[0038] In the formula, Let be the anomaly risk indicator for the j-th dynamic object in the target road at the t-th unit of time. Let be the velocity value of the j-th dynamic object in the target road at the t-th unit of time. Let be the acceleration value of the j-th dynamic object in the target road at the t-th unit of time. Let be the direction change value of the j-th dynamic object in the target road at the t-th unit of time. , , These represent the average values ​​of velocity, acceleration, and direction change, respectively. , , The weighting coefficients are, in order: motion speed weighting coefficient, motion speed weighting coefficient, and direction change weighting coefficient.

[0039] Preferably, step S3 further includes:

[0040] Based on the abnormal risk indicators of dynamic objects in the initial target road, and according to the GAM generalized additive model, the interaction effect between the abnormal risk indicators of each independent dynamic object in the initial target road is verified, and the interaction causal matrix of the abnormal risk indicators of dynamic objects in the initial target road is established.

[0041] Based on covariance analysis, the basic abnormal risk index of each dynamic object in the interaction causal matrix of the dynamic objects in the initial target road and the abnormal risk index of the dynamic objects under the interaction effect are verified.

[0042] The weighted evaluation formula uses the abnormal risk index of dynamic objects under the interaction effect as the basic abnormal risk index of each dynamic object in the interaction causal matrix between the collaborative risk factor and the abnormal risk index of dynamic objects in the target road, and calculates the abnormal risk index of dynamic objects in the target road.

[0043] Furthermore, a traffic video anomaly control system based on a 3D real-scene map is proposed to implement the traffic video anomaly control method based on a 3D real-scene map as described above, including:

[0044] The 3D modeling module is used to acquire traffic monitoring video streams and 3D modeling data of the target road based on heterogeneous traffic sensors, and to establish a 3D real-scene model of the target road.

[0045] The anomaly marking module is electrically connected to the 3D modeling module. The anomaly marking module is used to generate a spatial mapping between the traffic monitoring video stream and the 3D real-scene model of the target road based on the traffic monitoring video stream and the 3D real-scene model of the target road, and to mark the abnormal event information of dynamic objects in the target road.

[0046] The abnormal risk quantification module is electrically connected to the abnormal marking module. The abnormal risk quantification module is used to perform real-time three-dimensional rendering based on the abnormal event information of dynamic objects in the target road, generate a traffic state model of the three-dimensional real scene of the dynamic target road, and analyze the abnormal risk index of dynamic objects in the target road.

[0047] The control scheme generation module is electrically connected to the abnormal risk quantification module. The control scheme generation module is used to perform 3D real-scene modeling and visualization annotation of the target road based on the abnormal risk indicators of dynamic objects in the target road, determine the abnormal event type of the dynamic object, and generate a traffic abnormality control scheme for the target road.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] This invention proposes a traffic video anomaly control scheme based on a 3D real-scene map. By integrating heterogeneous traffic sensor data, it acquires traffic monitoring video streams and 3D modeling data for the target road, constructing a 3D real-scene model of the target road. Combining the spatial mapping between the traffic monitoring video stream and the 3D model, it marks abnormal target behavior data in real time and generates a dynamic traffic state model through 3D rendering, analyzing the risk indicators of abnormal targets. Finally, it visualizes and labels the types of abnormal events, generating targeted traffic anomaly solutions. The beneficial effects of this scheme are: improved traffic management efficiency, reduced probability of accidents, and optimized traffic flow management and emergency response. Attached Figure Description

[0050] Figure 1 This is a flowchart of a traffic video anomaly control method based on a 3D real-scene map;

[0051] Figure 2 This is a framework diagram of a traffic video anomaly control system based on a 3D real-scene map. Detailed Implementation

[0052] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0053] Reference Figure 1 As shown, a traffic video anomaly control method based on a 3D real-scene map includes:

[0054] S1. Based on heterogeneous traffic sensors, acquire traffic monitoring video streams and 3D modeling data of the target road, and establish a 3D real-scene model of the target road; the 3D modeling data includes: point cloud data, satellite imagery, and street view data;

[0055] Step S1 specifically includes:

[0056] Based on the multi-source data of traffic monitoring 3D modeling data of the target road, the ICP iterative nearest point algorithm is used to register the point cloud data to eliminate the bias, and satellite imagery and street view data are used as the texture source for 3D modeling to obtain the initial target road 3D real scene modeling data package.

[0057] Data preprocessing is performed on the initial 3D reality modeling data package of the target road;

[0058] As a further point, the ICP algorithm needs to ensure sufficient initial matching accuracy when registering point cloud data, otherwise it may lead to error accumulation; when satellite imagery and street view data are used as texture sources, it is necessary to ensure accurate mapping between texture and 3D model, and when using GDAL for coordinate transformation, it is necessary to ensure that the coordinate systems and projections of different data sources are consistent to prevent deviations during the transformation process; for the Poisson surface reconstruction algorithm, it is required that the normal vector estimation of point cloud data and sufficient point cloud density ensure that the generated target road 3D real scene model is smooth.

[0059] Using the GDAL data conversion open-source library, coordinate unification is performed on the traffic monitoring video stream of the target road and the initial 3D real scene modeling data package of the target road to obtain multi-source fused 3D real scene modeling data of the target road;

[0060] Based on multi-source fusion of target road 3D real scene modeling data, the Poisson surface reconstruction algorithm is used to solve the gradient field of the implicit function, which is consistent with the direction of the normal vector field of the input point cloud data, to generate the target road 3D real scene model.

[0061] Specifically, the ICP iterative nearest point algorithm is as follows:

[0062] In the formula, Let be the transformation matrix. To transform the matrix The point obtained by transforming the i-th source point. For the target point cloud There are 1 point, where m is the total amount of point cloud data.

[0063] The Poisson surface reconstruction algorithm is as follows:

[0064]

[0065] In the formula, For gradient operators, Implicit function At point gradient at, For point The normal vector at that location.

[0066] When using it, refer to the content in the steps above:

[0067] As a further step, fine registration using the ICP algorithm significantly reduces the stitching error of multi-source point cloud data, providing a high-precision data foundation for subsequent reconstruction. The Poisson reconstruction algorithm can reconstruct smooth and detailed surfaces from point clouds with normal vectors, and can infer partially occluded geometric structures. Coordinate unification using the GD library breaks down the barriers between data from different sensors, allowing 3D geometry (point cloud), aerial view information (satellite imagery), side view details (street view), and real-time dynamic streams (video) to be accurately superimposed on the same spatial reference. This provides a unique and reliable data base for subsequent spatial mapping and anomaly analysis. Using satellite imagery and street view data as texture sources, instead of using a single color or manual textures, ensures that the generated 3D model reproduces the real material, color, and appearance of the target road to the greatest extent. This not only enhances visual immersion but, more importantly, provides crucial real texture information for AI algorithms to identify road markings, traffic signs, vehicle models, etc. The output of this step is not a rough model, but a high-precision, highly realistic, multi-source fused 3D reality model. This high-quality model is an absolute prerequisite for achieving accurate mapping between video streams and three-dimensional space, and determines the accuracy and reliability of subsequent anomaly labeling and risk analysis.

[0068] S2. Based on the traffic monitoring video stream and the 3D real-scene model of the target road, generate a spatial mapping between the traffic monitoring video stream and the 3D real-scene model of the target road, and mark the abnormal event information of dynamic objects in the target road.

[0069] Step S2 specifically includes:

[0070] Based on the traffic monitoring video stream of the target road, according to the 3D real scene model of the target road, several sets of 2D and 3D feature image data point pairs are randomly selected from the traffic monitoring video stream of the target road and substituted into the cv2.solvePnP function to calculate and solve the camera extrinsic parameters, world coordinates and intrinsic parameters of the traffic monitoring video stream of the target road, and perform spatial mapping to obtain the spatial location information of the traffic monitoring video stream of the target road.

[0071] The traffic monitoring video stream of the target road is divided according to the frame rate of the video stream to obtain the traffic monitoring image frame sequence of the target road.

[0072] Step S2 also includes:

[0073] The spatial location information of the traffic monitoring video stream of the target road is used as the image attribute in the corresponding traffic monitoring image frame sequence of the target road. The YOLOv5 algorithm is used to collect the motion trajectory of dynamic objects in the traffic monitoring image frame sequence of the target road and extract the trajectory feature data of each dynamic object. The trajectory feature data includes: speed, acceleration, and direction change.

[0074] By using DeepSORT multi-target tracking, the trajectories of dynamic objects in adjacent frames of the trajectory feature data of each dynamic object are associated, and the position coordinates of dynamic objects in the complete data stream are marked to generate a sequence of motion object trajectory frames in the traffic monitoring image of the target road.

[0075] Step S2 also includes:

[0076] Based on the Lstm long short-term memory network, an abnormal target behavior trajectory analysis model is established. The target road traffic monitoring image moving object trajectory frame sequence is used as input. According to several trajectory step lengths of the target road traffic monitoring image moving object trajectory history, the future moving object trajectory is predicted, and abnormal target behavior label data in the target road is generated.

[0077] Based on the intrinsic and extrinsic parameters of traffic monitoring sensors on the target road, the two-dimensional spatial coordinates of abnormal target behavior label data in the target road are mapped to the three-dimensional real-scene model of the target road to obtain abnormal event information of dynamic objects in the target road.

[0078] Specifically, the target behavior trajectory anomaly analysis model is as follows:

[0079]

[0080] In the formula, This refers to the abnormal event information of the j-th dynamic object in the target road at the t-th unit of time. The first in the target road The hidden state of a dynamic object at time t. The first in the target road Trajectory feature data of a dynamic object at the t-th unit of time. The first in the target road Weight matrix of a dynamic object The first in the target road Bias terms for a dynamic object It is the sigmoid activation function.

[0081] When using it, refer to the content in the steps above:

[0082] As a further development, by establishing a spatial mapping between traffic monitoring videos and 3D reality models, the YOLOv5 and DeepSORT algorithms are used to accurately extract and track the motion trajectories of dynamic targets. Then, LSTM networks are combined to analyze the temporal features of the trajectories to intelligently identify abnormal behavior, ultimately mapping the abnormal events back to 3D space for localization. The beneficial effects of this solution are that it achieves a leap from two-dimensional planar perception to three-dimensional measurable space. This not only enables traffic monitoring to quantitatively analyze parameters such as target speed and distance, but also improves the intelligence and predictive foresight of abnormal event detection. Furthermore, the intuitive 3D visualization significantly enhances situational awareness and decision support capabilities, promoting the automation, precision, and efficiency of traffic management.

[0083] S3. Perform real-time 3D rendering based on the abnormal event information of dynamic objects in the target road, generate a traffic state model of the 3D real scene of the dynamic target road, and analyze the abnormal risk indicators of dynamic objects in the target road.

[0084] An observation window is established according to a unit time, and the abnormal event information of dynamic objects in the target road is used as the observation attribute to extract the abnormal behavior feature data of dynamic objects in the target road.

[0085] Based on the abnormal behavior characteristic data of dynamic objects in the target road, an abnormal behavior risk quantification function is established to calculate the abnormal risk index of dynamic objects in the initial target road.

[0086] Specifically, the abnormal behavior risk quantification function is as follows:

[0087]

[0088] In the formula, Let be the anomaly risk indicator for the j-th dynamic object in the target road at the t-th unit of time. Let be the velocity value of the j-th dynamic object in the target road at the t-th unit of time. Let be the acceleration value of the j-th dynamic object in the target road at the t-th unit of time. Let be the direction change value of the j-th dynamic object in the target road at the t-th unit of time. , , These represent the average values ​​of velocity, acceleration, and direction change, respectively. , , The weighting coefficients are, in order: motion speed weighting coefficient, motion speed weighting coefficient, and direction change weighting coefficient.

[0089] Step S3 also includes:

[0090] Based on the abnormal risk indicators of dynamic objects in the initial target road, and according to the GAM generalized additive model, the interaction effect between the abnormal risk indicators of each independent dynamic object in the initial target road is verified, and the interaction causal matrix of the abnormal risk indicators of dynamic objects in the initial target road is established.

[0091] Based on covariance analysis, the basic abnormal risk index of each dynamic object in the interaction causal matrix of the dynamic objects in the initial target road and the abnormal risk index of the dynamic objects under the interaction effect are verified.

[0092] The weighted evaluation formula uses the abnormal risk index of dynamic objects under the interaction effect as the basic abnormal risk index of each dynamic object in the interaction causal matrix between the collaborative risk factor and the abnormal risk index of dynamic objects in the target road, and calculates the abnormal risk index of dynamic objects in the target road.

[0093] S4. Based on the abnormal risk indicators of dynamic objects in the target road, perform 3D real-scene modeling and visualization annotation of the target road, determine the abnormal event types of dynamic objects, and generate a traffic anomaly control plan for the target road.

[0094] It should be noted that, based on the type of abnormal event, a traffic anomaly plan for the target road is generated. For example, the emergency response plan, based on the nature of the abnormal event, initiates corresponding emergency response measures, including on-site investigation, traffic diversion, and resource allocation. Simultaneously, control strategies are dynamically adjusted, optimizing traffic signals, adjusting traffic flow, and setting up detour routes in real time according to the development of the event. Early warnings and announcements are promptly issued through the information platform to ensure rapid restoration of order and minimize the impact on traffic.

[0095] When using it, please refer to the steps outlined above:

[0096] As a further development, this solution combines traffic monitoring video streams with 3D real-world modeling of the target road. It employs the Zhang Zhengyou calibration method to obtain the intrinsic and extrinsic parameters of the traffic monitoring sensors, extracts the motion trajectory features (such as velocity, acceleration, and direction changes) of dynamic objects based on the Yolov5 algorithm, and uses an LSTM network to analyze the abnormal behavior of these trajectory features. Finally, based on the sensor parameters, abnormal behavior labels in 2D space are mapped to the 3D real-world model of the target road, enabling real-time detection and labeling of abnormal events in dynamic objects. The beneficial effects of this solution are: it enables multi-target road monitoring, abnormal behavior prediction, and precise location, thereby improving traffic management and safety.

[0097] Reference Figure 2 As shown, based on the same inventive concept as a traffic video anomaly control method based on a 3D real-scene map, a traffic video anomaly control system based on a 3D real-scene map is proposed, including:

[0098] The 3D modeling module is used to acquire traffic monitoring video streams and 3D modeling data of the target road based on heterogeneous traffic sensors, and to establish a 3D real-scene model of the target road.

[0099] The anomaly marking module is electrically connected to the 3D modeling module. The anomaly marking module is used to generate a spatial mapping between the traffic monitoring video stream and the 3D real-scene model of the target road based on the traffic monitoring video stream and the 3D real-scene model of the target road, and to mark the abnormal event information of dynamic objects in the target road.

[0100] The abnormal risk quantification module is electrically connected to the abnormal marking module. The abnormal risk quantification module is used to perform real-time three-dimensional rendering based on the abnormal event information of dynamic objects in the target road, generate a traffic state model of the three-dimensional real scene of the dynamic target road, and analyze the abnormal risk index of dynamic objects in the target road.

[0101] The control scheme generation module is electrically connected to the abnormal risk quantification module. The control scheme generation module is used to perform 3D real-scene modeling and visualization annotation of the target road based on the abnormal risk indicators of dynamic objects in the target road, determine the abnormal event type of the dynamic object, and generate a traffic abnormality control scheme for the target road.

[0102] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for traffic video anomaly control based on a 3D real-scene map, characterized in that, include: S1. Based on heterogeneous traffic sensors, acquire traffic monitoring video streams and 3D modeling data of the target road, and establish a 3D real-scene model of the target road; The 3D modeling data includes: point cloud data, satellite imagery, and street view data. Step S1 specifically includes: Based on the multi-source data of traffic monitoring 3D modeling data of the target road, the ICP iterative nearest point algorithm is used to register the point cloud data to eliminate the bias, and satellite imagery and street view data are used as the texture source for 3D modeling to obtain the initial target road 3D real scene modeling data package. Data preprocessing is performed on the initial 3D reality modeling data package of the target road; Using the GDAL data conversion open-source library, coordinate unification is performed on the traffic monitoring video stream of the target road and the initial 3D real scene modeling data package of the target road to obtain multi-source fused 3D real scene modeling data of the target road; Based on multi-source fusion of target road 3D real scene modeling data, the Poisson surface reconstruction algorithm is used to solve the gradient field of the implicit function, which is consistent with the direction of the normal vector field of the input point cloud data, to generate the target road 3D real scene model. Specifically, the ICP iterative nearest point algorithm is as follows: ; In the formula, The transformation matrix is... To transform the matrix The point obtained by transforming the i-th source point. For the target point cloud There are 1 point, where m is the total amount of point cloud data; The Poisson surface reconstruction algorithm is as follows: ； In the formula, For gradient operators, implicit function At point gradient at, For point The normal vector at that location; S2. Based on the traffic monitoring video stream and the 3D real-scene model of the target road, generate a spatial mapping between the traffic monitoring video stream and the 3D real-scene model of the target road, and mark the abnormal event information of dynamic objects in the target road. S3. Based on the abnormal event information of dynamic objects in the target road, perform real-time 3D rendering to generate a traffic state model of the dynamic target road in 3D reality, and analyze the abnormal risk indicators of dynamic objects in the target road. Step S3 specifically includes: An observation window is established according to a unit time, and the abnormal event information of dynamic objects in the target road is used as the observation attribute to extract the abnormal behavior feature data of dynamic objects in the target road. Based on the abnormal behavior characteristic data of dynamic objects in the target road, an abnormal behavior risk quantification function is established to calculate the abnormal risk index of dynamic objects in the initial target road. Specifically, the abnormal behavior risk quantification function is as follows: ； In the formula, Let be the anomaly risk indicator for the j-th dynamic object in the target road at the t-th unit of time. Let be the velocity value of the j-th dynamic object in the target road at the t-th unit of time. Let be the acceleration value of the j-th dynamic object in the target road at the t-th unit of time. Let be the direction change value of the j-th dynamic object in the target road at the t-th unit of time. , , These represent the average values ​​of velocity, acceleration, and direction change, respectively. , , In order, they are the motion speed weighting coefficient, the motion speed weighting coefficient, and the direction change weighting coefficient; Step S3 also includes: Based on the abnormal risk indicators of dynamic objects in the initial target road, and according to the GAM generalized additive model, the interaction effect between the abnormal risk indicators of each independent dynamic object in the initial target road is verified, and the interaction causal matrix of the abnormal risk indicators of dynamic objects in the initial target road is established. Based on covariance analysis, the basic abnormal risk index of each dynamic object in the interaction causal matrix of the dynamic objects in the initial target road and the abnormal risk index of the dynamic objects under the interaction effect are verified. The weighted evaluation formula uses the abnormal risk index of dynamic objects under the interaction effect as the basic abnormal risk index of each dynamic object in the interaction causal matrix between the collaborative risk factor and the abnormal risk index of dynamic objects in the target road, and calculates the abnormal risk index of dynamic objects in the target road. S4. Based on the abnormal risk indicators of dynamic objects in the target road, perform 3D real-scene modeling and visualization annotation of the target road, determine the abnormal event types of dynamic objects, and generate a traffic anomaly control plan for the target road.

2. The traffic video anomaly control method based on a three-dimensional real-scene map according to claim 1, characterized in that, Step S2 specifically includes: Based on the traffic monitoring video stream of the target road, according to the 3D real scene model of the target road, several sets of 2D and 3D feature image data point pairs are randomly selected from the traffic monitoring video stream of the target road and substituted into the cv2.solvePnP function to calculate and solve the camera extrinsic parameters, world coordinates and intrinsic parameters of the traffic monitoring video stream of the target road, and perform spatial mapping to obtain the spatial location information of the traffic monitoring video stream of the target road. The traffic monitoring video stream of the target road is divided according to the frame rate of the video stream to obtain the traffic monitoring image frame sequence of the target road.

3. The traffic video anomaly control method based on a three-dimensional real-scene map according to claim 2, characterized in that, Step S2 also includes: The spatial location information of the traffic monitoring video stream of the target road is used as the image attribute in the corresponding traffic monitoring image frame sequence of the target road. The YOLOv5 algorithm is used to collect the motion trajectory of dynamic objects in the traffic monitoring image frame sequence of the target road and extract the trajectory feature data of each dynamic object. The trajectory feature data includes: speed, acceleration, and direction change. By using DeepSORT multi-target tracking, the trajectories of dynamic objects in adjacent frames of the trajectory feature data of each dynamic object are associated, and the position coordinates of dynamic objects in the complete data stream are marked to generate a sequence of motion object trajectory frames in the traffic monitoring image of the target road.

4. The traffic video anomaly control method based on a three-dimensional real-scene map according to claim 3, characterized in that, Step S2 also includes: Based on the Lstm long short-term memory network, an abnormal target behavior trajectory analysis model is established. The target road traffic monitoring image moving object trajectory frame sequence is used as input. According to several trajectory step lengths of the target road traffic monitoring image moving object trajectory history, the future moving object trajectory is predicted, and abnormal target behavior label data in the target road is generated. Based on the intrinsic and extrinsic parameters of traffic monitoring sensors on the target road, the two-dimensional spatial coordinates of abnormal target behavior label data in the target road are mapped to the three-dimensional real-scene model of the target road to obtain abnormal event information of dynamic objects in the target road. Specifically, the target behavior trajectory anomaly analysis model is as follows: ； In the formula, This refers to the abnormal event information of the j-th dynamic object in the target road at the t-th unit of time. The first in the target road The hidden state of a dynamic object at time t. The first in the target road Trajectory feature data of a dynamic object at the t-th unit of time. The first in the target road Weight matrix of a dynamic object The first in the target road Bias terms for a dynamic object It is the sigmoid activation function.

5. A traffic video anomaly control system based on a 3D real-scene map, characterized in that, A traffic video anomaly control method based on a three-dimensional real-scene map as described in any one of claims 1-4, comprising: The 3D modeling module is used to acquire traffic monitoring video streams and 3D modeling data of the target road based on heterogeneous traffic sensors, and to establish a 3D real-scene model of the target road. The anomaly marking module is electrically connected to the 3D modeling module. The anomaly marking module is used to generate a spatial mapping between the traffic monitoring video stream and the 3D real-scene model of the target road based on the traffic monitoring video stream and the 3D real-scene model of the target road, and to mark the abnormal event information of dynamic objects in the target road. The abnormal risk quantification module is electrically connected to the abnormal marking module. The abnormal risk quantification module is used to perform real-time three-dimensional rendering based on the abnormal event information of dynamic objects in the target road, generate a traffic state model of the three-dimensional real scene of the dynamic target road, and analyze the abnormal risk index of dynamic objects in the target road. The control scheme generation module is electrically connected to the abnormal risk quantification module. The control scheme generation module is used to perform 3D real-scene modeling and visualization annotation of the target road based on the abnormal risk indicators of dynamic objects in the target road, determine the abnormal event type of the dynamic object, and generate a traffic abnormality control scheme for the target road.

Images