Dam area traffic abnormal event monitoring system and method

Abstract

The invention discloses a dam area traffic abnormal event monitoring system and method, and belongs to the technical field of security and protection traffic monitoring, and the system comprises a vehicle track collection unit which collects video images of vehicles in a dam area and obtains dam area three-dimensional map data comprising geographic coordinates and semantic tags; the moving vehicle acquisition unit is used for detecting and separating a moving vehicle according to the video image and the three-dimensional map data; the abnormity identification unit is used for tracking the moving vehicle to form a formal path, and identifying overspeed, long-time stay or retrograde traffic abnormity events by analyzing the speed, the acceleration and the driving direction vector of the driving path; the abnormal event evaluation unit is used for grading the influence degree of the traffic abnormal event according to a preset multi-factor influence degree evaluation model; and the abnormity early warning unit is used for triggering video linkage early warning in a security monitoring center according to the grade of the influence degree. The problem that early warning cannot be effectively carried out on traffic abnormal events in time due to the fact that security and protection data are low in accuracy is solved.

Description

Technical Field

[0001] This invention relates to the field of security and traffic monitoring technology, specifically to a system and method for monitoring abnormal traffic events in dam areas. Background Technology

[0002] A security system is a system that utilizes various technologies such as audio-visual, infrared, detection, microwave, control, and communication, and employs various security products and equipment to provide a safe environment for people. The existing security monitoring system in the dam area mainly consists of three parts: front-end equipment, transmission network, and central equipment. Currently, the dam area security system still suffers from problems such as limited camera functionality and insufficient image clarity in some key areas, as well as monitoring blind spots. Traffic anomaly detection plays a crucial role in the dam area; traffic anomalies significantly reduce traffic efficiency, and the existing monitoring blind spots in the dam area security system prevent timely alarms and rescue efforts for traffic anomalies. Summary of the Invention

[0003] The purpose of this invention is to provide a monitoring system and method for abnormal traffic events in dam areas, addressing the aforementioned problems and solving the issue of the inability to provide timely and effective early warnings for abnormal events related to arrow sluices.

[0004] The technical solution of the present invention is as follows: A monitoring system for abnormal traffic events in dam areas, comprising: The vehicle trajectory acquisition unit collects video images of vehicles within the dam area and acquires 3D map data of the dam area, including geographic coordinates and semantic tags. The moving vehicle acquisition unit is connected to the vehicle trajectory acquisition unit. It detects and separates moving vehicles based on video images and 3D map data. Specifically, it spatially registers the 3D map data of the dam area with the video images to divide the static stable area and the dynamic interference area. It separates the foreground using an adaptive Gaussian mixture background model with a differentiated strategy, detects the vehicle bounding box using an improved YOLO model with a small target detection head, and performs a spatial logical AND operation on the foreground and vehicle bounding box to obtain the moving vehicle. The anomaly identification unit is connected to the moving vehicle acquisition unit. It tracks moving vehicles to form a path and identifies abnormal traffic events such as speeding, prolonged stopping, or driving in the wrong direction by analyzing the speed, acceleration, and direction vector of the driving path. The abnormal event evaluation unit, connected to the abnormality identification unit, classifies the degree of impact of traffic abnormal events according to a preset multi-factor impact evaluation model. The anomaly warning unit is connected to the anomaly event evaluation unit and triggers a video-linked warning at the security monitoring center based on the level of impact.

[0005] Furthermore, the improved YOLO model's network structure includes a tiny target detection head that integrates high-resolution shallow feature maps, and feature fusion and splicing operations are performed through upsampling and cross-layer splicing. During the model training phase, the improved YOLO model adopts a weighted loss function, with the loss component using FocalLoss: , in, The predicted probability of the model correctly classifying the sample is given by the following formula: ; in, The true category label of the sample; Predicting the category to which a sample belongs for the model The probability of; The weighting factor, used to balance the importance of positive and negative samples and adjust their contribution to the total loss, is formulated as follows: ,in ; To focus parameters, This is used to reduce the weight of easily distinguishable samples, so that the model training process focuses on difficult-to-distinguish samples. Improve the total loss function of the YOLO model Classification loss Bounding box regression loss and confidence loss Weighted sum: ; in, Used to measure the deviation between the predicted bounding box and the true bounding box. Used to measure the network's confidence in the existence of a target. These are the weighting coefficients for each loss term.

[0006] Furthermore, the adaptive Gaussian mixture background model separates the foreground through the following steps: For each pixel, the adaptive Gaussian mixture background model includes A Gaussian distribution, in At time 1, when a new pixel value is input, the model's weights are updated as follows: , in, It is the first A Gaussian distribution in Weight of time, It's the learning rate. The learning rate is 1 for a matching Gaussian distribution and 0 otherwise; Select based on the region to which the pixel belongs: If the pixel is located in a statically stable region, then If the pixel is located in the dynamic interference area, then ,and .

[0007] Furthermore, after improving the YOLO model and outputting the initial detection results, the mobile vehicle acquisition unit performs scene-aware post-processing operations; and before non-maximum suppression processing, it assigns a confidence score to each detection box. Make dynamic adjustments to obtain the adjusted score. : ; in, Based on the semantic labels of the region where the center point of the detection box is located in the 3D map The weighting function; then the adjusted score Perform NMS filtering; finally, perform a spatial logical AND operation on the foreground separated from the background model and the post-processed vehicle bounding boxes, and identify the vehicles corresponding to the vehicle bounding boxes whose foreground pixel ratio exceeds the preset area threshold as moving vehicles and output their position information.

[0008] Furthermore, the anomaly detection unit initializes a Kalman filter for each detected moving vehicle, and the filter's state vector... for ,in The location of the vehicle's center of gravity. The velocity component is used; the anomaly detection unit iteratively evaluates and predicts the vehicle's trajectory through two steps: prediction and update. The real-time speed of the vehicle is calculated based on the filtered state vector. and the direction vector; Traffic anomalies are identified based on the following criteria: real-time speed is compared with the speed limit value of the road segment obtained from 3D map data; exceeding the speed limit is considered speeding; whether the vehicle is continuously stationary in a non-parking area is determined; if the stationary duration exceeds a preset dwell time threshold, it is considered prolonged dwelling; and the vehicle's driving direction vector is compared with the standard driving direction vector of the lane obtained from 3D map data; if the angle between the two is greater than a preset angle threshold, it is considered driving in the wrong direction.

[0009] Furthermore, the abnormal event evaluation unit obtains the raw values ​​of evaluation indicators by analyzing vehicle behavior and traffic conditions in subsequent video frames; among them, the traffic congestion length is calculated by identifying continuous platoons with speeds below a preset threshold and measuring their length along the lane lines; the traffic flow reduction rate is calculated by setting up virtual loops downstream of the event point and comparing the difference between the number of vehicles passing through per unit time after the event and the historical average number of vehicles passing through during the same period; obtaining the raw values. Then, the dimensionless processing was performed using the min-max normalization method to obtain the normalized index value.

[0010] ; in: This refers to the original value of the indicator; and These are the minimum and maximum values ​​of the indicator within the statistical period, respectively; the normalized indicator values ​​are... Multiply by its corresponding weighting coefficient, which is preset according to the dam area safety management regulations. Then sum them up to get a comprehensive impact score. in: The total number of evaluation indicators; the comprehensive impact score will be calculated. The impact of the abnormal event is classified into mild, moderate, or severe levels by comparing it with multiple preset severity thresholds.

[0011] Furthermore, the anomaly warning unit receives the impact level output by the anomaly event evaluation unit; based on the impact level, the anomaly warning unit performs a graded warning operation at the security monitoring center and generates a structured data packet including event ID, timestamp, event type, location coordinates, severity level, and related video stream segment indexes, and pushes it to the upper-level management platform; for minor events, an icon flashes on the monitoring map to provide a notification; for moderate events, the relevant video footage automatically pops up; for severe events, in addition to popping up the video footage, an audible and visual alarm is triggered, and information such as event type, location, and impact level is overlaid and displayed on the video footage.

[0012] Furthermore, the system also includes: The checkpoint recording module provides a vehicle blacklist, which also includes images of the vehicle's historical violations. The dam area vehicle interception unit compares the currently entering vehicle with the vehicle blacklist, and supports image-based vehicle search in the blacklist based on the current vehicle image to prevent vehicles on the blacklist from entering the dam area. The vehicle access control unit, after confirming that the currently entering vehicle is not on the vehicle blacklist, records the vehicle's license plate number using license plate recognition technology and automatically grants the vehicle temporary access permission with a preset validity period.

[0013] This application also includes a method for monitoring abnormal traffic events in dam areas, applied to a system for monitoring abnormal traffic events in dam areas, comprising the following steps: Video images of vehicles within the dam area were collected, and 3D map data of the dam area containing geographic coordinates and semantic tags was obtained. The video images are registered with the 3D map data of the dam area to divide the static stable area and the dynamic interference area. An adaptive Gaussian mixture background model with a differentiated update strategy is used to separate the foreground. At the same time, an improved YOLO model with a small target detection head is used to detect the vehicle bounding boxes. The detection results are combined with the 3D map data for scene-aware post-processing. Finally, the moving vehicle is obtained by performing a spatial logical AND operation on the foreground and the post-processed vehicle bounding boxes. The Kalman filter algorithm is used to track the moving vehicle to form a driving path, and by analyzing the speed, acceleration and driving direction vector of the driving path, abnormal traffic events such as speeding, long-term stopping or driving in the wrong direction are identified. The length of traffic congestion, traffic flow, average vehicle speed and delay time caused by traffic anomalies are normalized and weighted and summed to classify the degree of impact of traffic anomalies. Based on the level of impact, a linked alert containing real-time video footage of the event will be triggered.

[0014] Furthermore, the improved YOLO model's network structure includes a tiny target detection head that integrates high-resolution shallow feature maps, and feature fusion and splicing operations are performed through upsampling and cross-layer splicing. During the model training phase, the improved YOLO model adopts a weighted loss function, with the loss component using FocalLoss: , in, The predicted probability of the model correctly classifying the sample is given by the following formula: ; in, The true category label of the sample; Predicting the category to which a sample belongs for the model The probability of; The weighting factor, used to balance the importance of positive and negative samples and adjust their contribution to the total loss, is formulated as follows: ,in ; To focus parameters, This is used to reduce the weight of easily distinguishable samples, so that the model training process focuses on difficult-to-distinguish samples. Improve the total loss function of the YOLO model Classification loss Bounding box regression loss and confidence loss Weighted sum: ; in, Used to measure the deviation between the predicted bounding box and the true bounding box. Used to measure the network's confidence in the existence of a target. These are the weighting coefficients for each loss term; The adaptive Gaussian mixture background model separates the foreground through the following steps: For each pixel, the adaptive Gaussian mixture background model includes A Gaussian distribution, in At time 1, when a new pixel value is input, the model's weights are updated as follows: , in, It is the first A Gaussian distribution in Weight of time, It's the learning rate. The learning rate is 1 for a matching Gaussian distribution and 0 otherwise; Select based on the region to which the pixel belongs: If the pixel is located in a statically stable region, then If the pixel is located in the dynamic interference area, then ,and .

[0015] Compared with existing technologies, the advantages of this invention are: By registering 3D map data of the dam area with video images, static stable zones and dynamic interference zones are delineated. A differentiated update strategy is applied to the adaptive Gaussian mixture background model to specifically suppress interference from unique environmental factors of the dam area, such as water surface reflection and swaying vegetation, thereby improving the accuracy of moving vehicle detection. A multi-factor influence evaluation model is used to quantitatively assess and classify abnormal traffic events, enabling precise early warning.

[0016] The system integrates blacklist management and vehicle access control, enhancing the proactiveness and efficiency of safety management in the dam area. The overall system improves the accuracy of security data, ensuring timely and effective early warning and response to traffic anomalies. Attached Figure Description

[0017] Figure 1 This is a diagram showing the overall structural framework of the system in this application.

[0018] Figure 2This is the overall flowchart of the method in this application.

[0019] Figure 3 This is a flowchart of the steps for obtaining the mobile vehicle in this application.

[0020] Figure 4 This is a schematic diagram illustrating the registration of video images with a 3D map and the division of regions in this application.

[0021] Figure 5 This is a schematic diagram of the improved YOLO model network structure in this application.

[0022] Figure 6 This is a schematic diagram illustrating the vehicle tracking and abnormal event identification of this application.

[0023] Figure 7 This is a schematic diagram of the abnormal event impact assessment and graded early warning interface in this application.

[0024] Attached reference numerals: 100-Vehicle trajectory acquisition unit, 200-Moving vehicle acquisition unit, 300-Anomaly identification unit, 400-Anomaly event evaluation unit, 500-Anomaly early warning unit, 600-Clipping recording module, 700-Dam area vehicle interception unit, 800-Vehicle permission setting unit. Detailed Implementation

[0025] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0026] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0027] Please see Figure 1-7 A monitoring system for abnormal traffic events in dam areas, such as Figure 1 As shown, it includes a vehicle trajectory acquisition unit 100, a moving vehicle acquisition unit 200, an anomaly identification unit 300, an anomaly event evaluation unit 400, and an anomaly early warning unit 500.

[0028] The system also includes a checkpoint recording module 600 for dam area entrance management, a dam area vehicle interception unit 700, and a vehicle permission setting unit 800. The checkpoint recording module 600 is a database system used to structurally store vehicle blacklist information, including license plate number, vehicle color, vehicle model, historical violation images and their corresponding feature vectors, violation type, and timestamp. The dam area vehicle interception unit 700 consists of a high-definition license plate recognition camera, a vehicle panoramic camera, a local edge computing processor, and a physical barrier gate. When a vehicle enters, the license plate recognition camera and the panoramic camera respectively capture the license plate image and the vehicle panoramic image.

[0029] The local edge computing processor first parses the license plate number string from the license plate image and uses it as an index to initiate a first-stage query to the checkpoint recording module 600. If the query matches, the vehicle is identified as a blacklisted vehicle, and the processor immediately sends a control signal to the physical barrier to maintain its closed state. If the first-stage query fails, or if the license plate number cannot be parsed due to damage or obstruction, the processor extracts the vehicle's deep learning feature vector from the panoramic image and initiates a second-stage query based on feature vector comparison to the checkpoint recording module 600. This query calculates the cosine similarity between the input vehicle feature vector and the feature vectors of historical violation images stored in the blacklist. If the similarity is higher than a preset threshold, the vehicle is also identified as a blacklisted vehicle and interception is executed.

[0030] If neither of the two-stage queries finds a match, the vehicle is determined to be authorized. At this point, the vehicle access control unit 800 is activated, recording the vehicle's license plate number, panoramic image, and entry timestamp, and generating a temporary pass with a preset validity period (e.g., 24 hours) and storing it in the access log database. Simultaneously, the processor of the dam area vehicle interception unit 700 sends an opening signal to the physical barrier gate, granting the vehicle entry.

[0031] The vehicle trajectory acquisition unit 100 consists of high-definition road imaging equipment deployed on key road sections in the dam area, optional infrared thermal imaging equipment, and a 3D map data server. The high-definition equipment is used to acquire video under normal lighting conditions, while the infrared thermal imaging equipment is used to acquire video under low visibility conditions such as nighttime or foggy weather. This unit ensures the synchronization of timestamps for all video streams via Network Time Protocol (NTP). The 3D map data server provides high-precision 3D map data, which includes geometric information such as road centerlines, lane boundaries, speed limit signs, slopes, and curvature, as well as semantic labels for areas such as road surfaces, water surfaces, and vegetation.

[0032] like Figure 3 and Figure 4As shown, the mobile vehicle acquisition unit 200 is connected to the vehicle trajectory acquisition unit 100 and receives video images and 3D map data. This unit first establishes a precise mapping relationship between the 3D map world coordinate system and the 2D video image pixel coordinate system using camera calibration parameters (including intrinsic and extrinsic parameter matrices) and the PnP (Perspective-n-Point) algorithm, achieving spatial registration. Based on this mapping, each pixel in the video frame is accurately classified into either a static stable region or a dynamic interference region.

[0033] Subsequently, the moving vehicle acquisition unit 200 uses an adaptive Gaussian mixture background model to model the background of the video sequence. For each pixel, its background model is... Composed of Gaussian distributions, in At any given moment, when a new pixel value is input, the model's weights are updated according to the following rules: ; in: It is the first A Gaussian distribution in The weight of each moment; It is the learning rate; The learning rate is 1 for a matching Gaussian distribution and 0 otherwise. Select based on the region to which the pixel belongs: If the pixel is located in a statically stable region, then (A relatively large value); if the pixel is located in a dynamic interference area, then (a smaller value, and) ).

[0034] Simultaneously, the mobile vehicle acquisition unit 200 performs vehicle target detection. Before inputting the video image into the vehicle detection model, a lightweight region nomination network locates potential tiny target regions with pixel sizes smaller than a preset size threshold, and only performs bicubic interpolation on image patches in these regions for local magnification. The processed image is then input into an improved YOLO model for vehicle detection. This improved YOLO model incorporates a dedicated tiny target detection head that fuses high-resolution shallow feature maps from earlier layers (such as layer C2) in the network backbone. Feature fusion is achieved through upsampling and cross-layer concatenation operations.

[0035] like Figure 5 As shown, during the model training phase, this improved YOLO model employs a weighted loss function, with the classification loss using FocalLoss, defined as: ; in: The predicted probability that the model correctly classifies the sample is defined as: ; The true category label of the sample; Predicting the category to which a sample belongs for the model The probability of; A weighting factor, used to balance the importance of positive and negative samples and adjust their contribution to the total loss, is defined as follows: ,in ; To focus parameters, This is used to reduce the weight of easily distinguishable samples, so that the model training process focuses on difficult-to-distinguish samples.

[0036] The total loss function of the model Classification loss Bounding box regression loss and confidence loss Weighted sum: ; in, Used to measure the deviation between the predicted bounding box and the true bounding box. Used to measure the network's confidence in the existence of a target. These are the weighting coefficients for each loss term. This loss function introduces FocalLoss to reduce the weight of easily distinguishable background samples, while simultaneously increasing the weight of the detection and classification losses for small target samples in the total loss function.

[0037] After the model outputs the initial detection results, the moving vehicle acquisition unit 200 performs a scene-aware post-processing operation.

[0038] After the model outputs the initial detection results, the moving vehicle acquisition unit 200 performs a scene-aware post-processing operation. Before non-maximum suppression (NMS) processing, a confidence score is calculated for each detection box. Make dynamic adjustments to obtain the adjusted score. : ; in, Based on the semantic labels of the region where the center point of the detection box is located in the 3D map The weight function, for example, when When it is "road" ,when When it is "water area" Then adjust the scores. Perform NMS filtering. Finally, perform a spatial logical AND operation between the foreground separated from the background model and the post-processed vehicle bounding boxes, and identify vehicles whose foreground pixel ratio exceeds a preset area threshold as moving vehicles and output their position information.

[0039] like Figure 6 As shown, the anomaly detection unit 300 receives the location information of moving vehicles. This unit initializes a Kalman filter for each detected moving vehicle. The filter's state vector... Defined as ,in The location of the vehicle's center of gravity. The velocity component is used. This unit iterates through two steps—prediction and update—to evaluate and predict the vehicle's trajectory. Based on the filtered state vector, the real-time velocity of the vehicle is calculated. And the driving direction vector. Then, traffic anomalies are identified according to the following criteria: the real-time speed is compared with the speed limit value of the road segment obtained from the 3D map data, and exceeding the speed limit is judged as speeding; it is determined whether the vehicle is continuously stationary in a non-parking area, and if the stationary time exceeds the preset dwell time threshold, it is judged as long-term dwelling; the vehicle's driving direction vector is compared with the standard driving direction vector of the lane obtained from the 3D map data, and if the angle between the two is greater than the preset angle threshold, it is judged as driving in the wrong direction.

[0040] like Figure 7 As shown, the abnormal event evaluation unit 400 is activated after the abnormal event identification unit 300 identifies a traffic anomaly. This unit obtains the raw values ​​of evaluation indicators by analyzing vehicle behavior and traffic conditions in subsequent video frames. For example, traffic congestion length is calculated by identifying continuous platoons with speeds below a preset threshold and measuring their length along the lane lines; traffic flow reduction rate is calculated by setting up virtual loops downstream of the event point and comparing the difference between the number of vehicles passing through per unit time after the event and the historical average number of vehicles passing through during the same period. (Original values ​​are obtained.) Then, the dimensionless processing was performed using the min-max normalization method to obtain the normalized index value.

[0041] ; in: This refers to the original value of the indicator; and These are the minimum and maximum values ​​of the indicator within the statistical period, respectively. Then, the normalized indicator values ​​are... Multiply by its corresponding weighting coefficient, which is preset according to the dam area safety management regulations. Then sum them up to get a comprehensive impact score. in: This represents the total number of evaluation indicators. Finally, the overall impact score will be determined. The impact of the abnormal event is classified into mild, moderate, or severe levels by comparing it with multiple preset severity thresholds.

[0042] The anomaly warning unit 500 receives the impact level output from the anomaly event evaluation unit 400. Based on this level, the unit performs a graded warning operation at the security monitoring center and generates a structured data packet containing the event ID, timestamp, event type, location coordinates, severity level, and related video stream segment indexes, which is then pushed to the upper-level management platform. For minor events, an icon flashes on the monitoring map; for moderate events, relevant video footage automatically pops up; for severe events, in addition to popping up the video footage, an audible and visual alarm is triggered, and information such as the event type, location, and impact level is overlaid on the video footage.

[0043] like Figure 2 As shown, this application also provides a method for monitoring abnormal traffic events in dam areas, including the following steps: S1. Collect video images of vehicles within the dam area and obtain 3D map data of the dam area containing geographic coordinates and semantic tags; This step involves activating high-definition and optional infrared road surveillance equipment deployed along key sections of the dam area to continuously capture video streams at a preset frame rate (e.g., 25 frames per second), and using the Network Time Protocol (NTP) to precisely timestamp and synchronize all video frames. Simultaneously, based on the camera's geographical coordinates, the system loads high-precision 3D map data for the corresponding area from a 3D map data server. This 3D map data not only includes the precise 3D spatial coordinates of objects such as roads, bridges, and water bodies, but also adds semantic labels to the surfaces of these objects.

[0044] S2. The video images are registered with the 3D map data of the dam area to divide the static stable area and the dynamic interference area. An adaptive Gaussian mixture background model with a differentiated update strategy is used to separate the foreground. At the same time, an improved YOLO model with a small target detection head is used to detect the vehicle bounding boxes. The detection results are combined with the 3D map data to perform scene-aware post-processing. Finally, the moving vehicles are obtained by performing a spatial logical AND operation on the foreground and the post-processed vehicle bounding boxes. This step first performs spatial registration: using a pre-calibrated matrix of camera intrinsic and extrinsic parameters, the Perspective-n-Point (PnP) algorithm is used to establish a mapping between the 3D map world coordinate system and the 2D video image pixel coordinate system for each video frame. Based on this mapping, each pixel in the current frame is precisely classified into either a statically stable region or a dynamically disturbed region.

[0045] Subsequently, background modeling and object detection are performed in parallel. In background modeling, an adaptive Gaussian mixture model is used. When updating this model, a higher initial learning rate is applied to pixels classified as statically stable regions. To quickly adapt to global changes such as lighting; a lower second update learning rate is used for pixels classified as dynamic interference areas. To suppress model disturbances caused by local dynamic factors such as water surface reflection and vegetation swaying, a foreground binarized mask image is finally generated.

[0046] In object detection, the current video frame is first preprocessed by using a lightweight region nomination network to quickly locate potential tiny target regions with pixel sizes smaller than a preset threshold, and then performing bicubic interpolation for local magnification only on these regions. The processed image is then input into an improved YOLO model with a tiny object detection head. This model processes the image and outputs preliminary vehicle bounding boxes and confidence scores. Next, scene-aware post-processing is performed: for each bounding box, the semantic label of its center point in the 3D map is queried using the established mapping relationship, and its confidence score is dynamically adjusted based on the label (such as "road" or "water"). After adjustment, the non-maximum suppression (NMS) algorithm is executed to obtain a set of optimized final vehicle bounding boxes.

[0047] Finally, collaborative decision-making is performed to identify moving vehicles. A spatial logical AND operation is performed between the foreground binarized mask image and the final set of vehicle bounding boxes. For each bounding box, the area ratio of its internal foreground pixels is calculated, and vehicles whose ratio exceeds a preset area threshold are identified as moving vehicles.

[0048] S3. Use the Kalman filter algorithm to track moving vehicles to form a driving path, and identify traffic anomalies such as speeding, long-term stopping or driving in the wrong direction by analyzing the speed, acceleration and driving direction vector of the driving path. In this step, for each moving vehicle identified in the previous step, the system initializes or associates a Kalman filter instance. The state vector of this filter contains the two-dimensional position and velocity components of the vehicle's centroid. For each subsequent frame of the video stream, the vehicle state is optimally estimated using the Kalman filter's prediction and update mechanism. By continuously recording the smoothed centroid coordinates of the vehicle output by the filter in each frame, a precise spatiotemporal travel path is generated for each vehicle. Based on this path, the real-time velocity and direction vector of the vehicle are obtained by calculating the spatiotemporal displacement between path points. These real-time motion parameters are then compared in real-time with the road attributes corresponding to the vehicle's current location, retrieved from the 3D map data. When preset anomaly conditions are met, it is identified as a traffic anomaly event.

[0049] S4. Normalize and weight the length of traffic congestion, traffic flow, average vehicle speed and delay time caused by traffic anomalies to classify the degree of impact of traffic anomalies. This step is triggered once an abnormal event is identified. The system continuously analyzes the traffic conditions downstream of the event point, tracks subsequent vehicle trajectories, and calculates a series of preset evaluation indicators, such as quantifying traffic congestion length by identifying the length of continuous slow-moving vehicle queues. After obtaining the raw values ​​of each indicator, the min-max normalization method is used to convert them into dimensionless normalized values ​​within the interval. Then, according to the weighting coefficients preset for each indicator in the dam area safety management regulations, these normalized values ​​are weighted and summed to calculate a comprehensive impact score. Finally, this score is compared with preset severity level threshold intervals (e.g., [0,0.4] for mild, [0.4,0.7] for moderate, and [0.7,1.0] for severe) to objectively classify the degree of impact of the event into the corresponding level.

[0050] S5. Based on the level of impact, trigger a linked alert that includes real-time video footage of the event; This step executes differentiated early warning operations based on the impact level assessed in the previous step. If the level is mild, a flashing icon will be displayed on the electronic map of the security monitoring center. If the level is moderate, a real-time video feed of the event's location will automatically pop up on the main monitoring screen. If the level is severe, the audible and visual alarm devices in the control room will be triggered simultaneously with the video feed, and structured event details (including event type, location coordinates, impact level, etc.) will be prominently overlaid on the video feed and pushed to the upper-level emergency command platform.

[0051] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A monitoring system for abnormal traffic events in dam areas, characterized in that, include: The vehicle trajectory acquisition unit collects video images of vehicles within the dam area and acquires 3D map data of the dam area, including geographic coordinates and semantic tags. The moving vehicle acquisition unit is connected to the vehicle trajectory acquisition unit. It detects and separates moving vehicles based on video images and 3D map data. Specifically, it spatially registers the 3D map data of the dam area with the video images to divide the static stable area and the dynamic interference area. It separates the foreground using an adaptive Gaussian mixture background model with a differentiated strategy, detects the vehicle bounding box using an improved YOLO model with a small target detection head, and performs a spatial logical AND operation on the foreground and vehicle bounding box to obtain the moving vehicle. The anomaly identification unit is connected to the moving vehicle acquisition unit. It tracks moving vehicles to form a path and identifies abnormal traffic events such as speeding, prolonged stopping, or driving in the wrong direction by analyzing the speed, acceleration, and direction vector of the driving path. The abnormal event evaluation unit, connected to the abnormality identification unit, classifies the degree of impact of traffic abnormal events according to a preset multi-factor impact evaluation model. The anomaly warning unit is connected to the anomaly event evaluation unit and triggers a video-linked warning at the security monitoring center based on the level of impact.

2. The dam area traffic anomaly monitoring system according to claim 1, characterized in that, The improved YOLO model's network structure includes a tiny target detection head that integrates high-resolution shallow feature maps, and feature fusion and splicing operations are performed through upsampling and cross-layer splicing. During the model training phase, the improved YOLO model adopts a weighted loss function, with the loss component using FocalLoss: ， in, The predicted probability of the model correctly classifying the sample is given by the following formula: ； in, The true category label of the sample; Predicting the category to which a sample belongs for the model The probability of; The weighting factor, used to balance the importance of positive and negative samples and adjust their contribution to the total loss, is formulated as follows: ,in ; To focus parameters, This is used to reduce the weight of easily distinguishable samples, so that the model training process focuses on difficult-to-distinguish samples. Improve the total loss function of the YOLO model Classification loss Bounding box regression loss and confidence loss Weighted sum: ； in, Used to measure the deviation between the predicted bounding box and the true bounding box. Used to measure the network's confidence in the existence of a target. These are the weighting coefficients for each loss term.

3. The dam area traffic anomaly monitoring system according to claim 2, characterized in that, The adaptive Gaussian mixture background model separates the foreground through the following steps: For each pixel, the adaptive Gaussian mixture background model includes A Gaussian distribution, in At time 1, when a new pixel value is input, the model's weights are updated as follows: ， in, It is the first A Gaussian distribution in Weight of time, It's the learning rate. The learning rate is 1 for a matching Gaussian distribution and 0 otherwise; Select based on the region to which the pixel belongs: If the pixel is located in a statically stable region, then If the pixel is located in the dynamic interference area, then ,and .

4. The dam area traffic anomaly monitoring system according to claim 3, characterized in that, After the mobile vehicle acquisition unit outputs preliminary detection results using the improved YOLO model, it performs post-processing operations for scene awareness; and before non-maximum suppression processing, it assigns a confidence score to each detection box. Make dynamic adjustments to obtain the adjusted score. : ； in, Based on the semantic labels of the region where the center point of the detection box is located in the 3D map The weighting function; then the adjusted score Perform NMS filtering; finally, perform a spatial logical AND operation on the foreground separated from the background model and the post-processed vehicle bounding boxes, and identify the vehicles corresponding to the vehicle bounding boxes whose foreground pixel ratio exceeds the preset area threshold as moving vehicles and output their position information.

5. The dam area traffic anomaly monitoring system according to claim 1, characterized in that, The anomaly detection unit initializes a Kalman filter for each detected moving vehicle, and the filter's state vector... for ,in The location of the vehicle's center of gravity. The velocity component is used; the anomaly detection unit iteratively evaluates and predicts the vehicle's trajectory through two steps: prediction and update. The real-time speed of the vehicle is calculated based on the filtered state vector. and the direction vector; Traffic anomalies are identified according to the following criteria: real-time speed is compared with the speed limit value of the road segment obtained from 3D map data, and exceeding the speed limit value is determined as speeding; it is determined whether the vehicle is continuously stationary in a non-parking area, and if the stationary time exceeds the preset dwell time threshold, it is determined as long-term dwelling. The vehicle's driving direction vector is compared with the standard driving direction vector of its lane obtained from the 3D map data. If the angle between the two is greater than a preset angle threshold, it is determined to be a reverse driving behavior.

6. The dam area traffic anomaly monitoring system according to claim 5, characterized in that, The abnormal event evaluation unit obtains the raw values ​​of evaluation indicators by analyzing vehicle behavior and traffic conditions in subsequent video frames. Specifically, traffic congestion length is calculated by identifying continuous platoons with speeds below a preset threshold and measuring their length along the lane lines; traffic flow reduction rate is calculated by setting up virtual loops downstream of the event point and comparing the difference between the number of vehicles passing per unit time after the event and the historical average number of vehicles passing during the same period. Then, the dimensionless processing was performed using the min-max normalization method to obtain the normalized index value. ； in: This refers to the original value of the indicator; and These are the minimum and maximum values ​​of the indicator within the statistical period, respectively; the normalized indicator values ​​are... Multiply by its corresponding weighting coefficient, which is preset according to the dam area safety management regulations. Then sum them up to get a comprehensive impact score. in: The total number of evaluation indicators; the comprehensive impact score will be calculated. The impact of the abnormal event is classified into mild, moderate, or severe levels by comparing it with multiple preset severity thresholds.

7. A monitoring system for abnormal traffic events in a dam area according to claim 6, characterized in that, The anomaly warning unit receives the impact level output by the anomaly event evaluation unit; based on the impact level, the anomaly warning unit performs a graded warning operation in the security monitoring center and generates a structured data packet including event ID, timestamp, event type, location coordinates, severity level and related video stream segment index, and pushes it to the upper management platform; For minor incidents, an icon will flash on the monitoring map to indicate the incident; for moderate incidents, relevant video footage will automatically pop up. For serious incidents, in addition to displaying a video feed, an audible and visual alarm will be triggered, and information such as the incident type, location, and impact level will be overlaid on the video feed.

8. A monitoring system for abnormal traffic events in a dam area according to claim 1, characterized in that, The system also includes: The checkpoint recording module provides a vehicle blacklist, which also includes images of the vehicle's historical violations. The dam area vehicle interception unit compares the currently entering vehicle with the vehicle blacklist, and supports image-based vehicle search in the blacklist based on the current vehicle image to prevent vehicles on the blacklist from entering the dam area. The vehicle access control unit, after confirming that the currently entering vehicle is not on the vehicle blacklist, records the vehicle's license plate number using license plate recognition technology and automatically grants the vehicle temporary access permission with a preset validity period.

9. A method for monitoring abnormal traffic events in dam areas, characterized in that, The system for monitoring abnormal traffic events in a dam area, as described in any one of claims 1-8, comprises the following steps: Video images of vehicles within the dam area were collected, and 3D map data of the dam area containing geographic coordinates and semantic tags was obtained. The video images are registered with the 3D map data of the dam area to divide the static stable area and the dynamic interference area. An adaptive Gaussian mixture background model with a differentiated update strategy is used to separate the foreground. At the same time, an improved YOLO model with a small target detection head is used to detect the vehicle bounding boxes. The detection results are combined with the 3D map data for scene-aware post-processing. Finally, the moving vehicle is obtained by performing a spatial logical AND operation on the foreground and the post-processed vehicle bounding boxes. The Kalman filter algorithm is used to track the moving vehicle to form a driving path, and by analyzing the speed, acceleration and driving direction vector of the driving path, abnormal traffic events such as speeding, long-term stopping or driving in the wrong direction are identified. The length of traffic congestion, traffic flow, average vehicle speed and delay time caused by traffic anomalies are normalized and weighted and summed to classify the degree of impact of traffic anomalies. Based on the level of impact, a linked alert containing real-time video footage of the event will be triggered.

10. A method for monitoring abnormal traffic events in a dam area according to claim 9, characterized in that, The improved YOLO model's network structure includes a tiny target detection head that integrates high-resolution shallow feature maps, and feature fusion and splicing operations are performed through upsampling and cross-layer splicing. During the model training phase, the improved YOLO model adopts a weighted loss function, with the loss component using FocalLoss: ， in, The predicted probability of the model correctly classifying the sample is given by the following formula: ； in, The true category label of the sample; Predicting the category to which a sample belongs for the model The probability of; The weighting factor, used to balance the importance of positive and negative samples and adjust their contribution to the total loss, is formulated as follows: ,in ; To focus parameters, This is used to reduce the weight of easily distinguishable samples, so that the model training process focuses on difficult-to-distinguish samples. Improve the total loss function of the YOLO model Classification loss Bounding box regression loss and confidence loss Weighted sum: ； in, Used to measure the deviation between the predicted bounding box and the true bounding box. Used to measure the network's confidence in the existence of a target. These are the weighting coefficients for each loss term; The adaptive Gaussian mixture background model separates the foreground through the following steps: For each pixel, the adaptive Gaussian mixture background model includes A Gaussian distribution, in At time 1, when a new pixel value is input, the model's weights are updated as follows: ， in, It is the first A Gaussian distribution in Weight of time, It's the learning rate. The learning rate is 1 for a matching Gaussian distribution and 0 otherwise; Select based on the region to which the pixel belongs: If the pixel is located in a statically stable region, then If the pixel is located in the dynamic interference area, then ,and .

Images