A multi-camera cooperative event detection method for full coverage of expressway

Abstract

The application discloses a multi-camera cooperative event detection method for full-coverage laying of expressways and belongs to the technical field of intelligent transportation. The method is characterized in that a plurality of cameras are cross-laid along the roadside of an expressway to achieve full-domain video stream acquisition and structured processing of bidirectional lanes, target detection, feature extraction and single-camera tracking technology are used to identify foreground targets, and a cross-camera cooperative mechanism is used to perform relay tracking and behavior analysis on targets such as litter, pedestrians and vehicles, and finally, a risk assessment model based on multi-dimensional weighted fusion is used to output graded early warning information. The application realizes timely discovery, accurate identification and risk assessment of sudden abnormal events.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation technology, and in particular relates to a multi-camera collaborative event detection method for full coverage of highways. Background Technology

[0002] In recent years, with the continuous expansion of the expressway network and the sustained improvement of traffic efficiency, higher demands have been placed on the real-time detection and rapid response capabilities for sudden abnormal events on the road. In particular, in the monitoring of emergencies such as littering, pedestrians running into traffic, and illegal parking, failure to achieve timely detection and early warning can easily lead to secondary accidents and seriously threaten driving safety.

[0003] Currently, highway incident detection mainly relies on two types of technologies: one is detection methods based on multi-source data fusion, but these are usually limited by the deployment range of existing detection equipment, making it difficult to achieve full road coverage, and they mainly focus on traffic flow anomaly analysis, with limited ability to detect static or small target events; the other is detection schemes based on video data, which can achieve a certain range of visual coverage, but existing systems mainly monitor vehicle driving status, and are still insufficient in dealing with the identification of non-vehicle targets such as debris and pedestrians, as well as continuous tracking across cameras, and lack quantitative assessment and graded early warning mechanisms for incident risks.

[0004] Therefore, there is an urgent need to propose a detection method that can achieve seamless coverage of the entire highway, support accurate identification of multiple types of targets and cross-camera collaborative tracking, and intelligently assess and classify event risks, so as to improve the timeliness and effectiveness of handling abnormal events on highways. Summary of the Invention

[0005] This invention proposes a multi-camera collaborative event detection method for full coverage deployment on highways, in order to solve the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides a multi-camera collaborative event detection method for full coverage deployment on highways, comprising:

[0007] Multi-camera wide-area video acquisition devices are deployed intersecting along the target road segment to achieve full coverage video acquisition of the target road segment and obtain multiple video streams;

[0008] In the data analysis server, a camera processing unit is established for each camera to decode the video stream into image sequence frames, and each frame image, along with its acquisition timestamp and road segment number, is encapsulated into a raw data unit.

[0009] The image frames in the original data unit are subjected to ROI candidate region extraction to obtain the target candidate ROI region;

[0010] Target detection is performed on the candidate ROI region to obtain the target ROI coordinates, target type and confidence level, and target feature vector is extracted;

[0011] The target feature vector is input into the target tracking algorithm to update the target movement trajectory, and the target trajectory ID, target ROI coordinates, target type, confidence level, target feature vector and acquisition time are packaged into target detection data;

[0012] In each camera processing unit, local target tracking is performed based on the target detection data to establish a target tracking sequence;

[0013] The target tracking sequence is subjected to velocity trend analysis to obtain target velocity trend markers;

[0014] Based on the target velocity trend marker and target type, multi-camera collaborative tracking processing is performed;

[0015] Based on the collaborative tracking results, a risk assessment is performed on the target tracking sequence, and the risk level is output.

[0016] Optionally, the multi-camera wide-area video acquisition device includes a column and a camera group installed on the top of the column, the camera group including a PTZ camera, a short-focus bullet camera and a long-focus bullet camera;

[0017] Based on the deployment positions of the PTZ camera, short-focus camera, and long-focus camera, video data for different areas is acquired;

[0018] Based on the overlapping areas of the fields of view between adjacent cameras, obtain the cross-over video data.

[0019] Optionally, ROI candidate region extraction for the image frames in the original data unit includes:

[0020] Update the background image using the moving average background modeling method;

[0021] Obtain the difference image based on the difference operation between the current image frame and the background image;

[0022] Based on the binarization thresholding of the difference image, connected components are extracted as candidate ROI regions.

[0023] Optionally, target detection of the candidate ROI region includes:

[0024] Based on the pre-trained YOLOV10 model, the target candidate ROI region is detected, and the target ROI coordinates, target type and confidence score are obtained;

[0025] Based on the pre-trained ConvNeXt-L model, feature extraction is performed on the target ROI region to obtain the target feature vector.

[0026] Optionally, inputting the target feature vector into the target tracking algorithm includes:

[0027] Based on the DeepSORT algorithm, the target's trajectory is updated using the target's feature vector.

[0028] Target detection data is generated based on the target trajectory ID, target ROI coordinates, target type, confidence level, target feature vector, and acquisition time.

[0029] Optionally, local target tracking in each camera processing unit includes:

[0030] Based on the target trajectory ID in the target detection data, add the corresponding data to the existing target tracking sequence;

[0031] Based on the similarity between target feature vectors, unmatched target detection data are matched with existing target tracking sequences;

[0032] Update or create a new target tracking sequence based on the matching results.

[0033] Optionally, velocity trend analysis of the target tracking sequence includes:

[0034] Extract multiple target detection datasets based on time periods;

[0035] Based on the calibration relationship between pixels and actual distance, calculate the displacement distance and velocity of the target in each cycle;

[0036] Based on the speed change, determine whether the target is stationary, accelerating, or decelerating, and generate a speed trend marker.

[0037] Optionally, multi-camera cooperative tracking processing includes:

[0038] Based on the target type and confidence level, determine whether it is spilled material, a pedestrian, or a vehicle;

[0039] Based on the target velocity trend markers and the target tracking sequences of adjacent cameras, the projectile risk correction parameters are calculated;

[0040] Based on the situation where the target enters the overlapping area of ​​the field of view of adjacent cameras, cross-camera target tracking sequences are associated and stitched together.

[0041] Optionally, risk assessment of the target tracking sequence includes:

[0042] Based on the target type, calculate the risk score for pedestrians who break in, the risk score for illegally parked vehicles, and the risk score for debris;

[0043] Based on the comparison between the risk score and the preset threshold, the corresponding risk level warning is triggered.

[0044] Compared with the prior art, the present invention has the following advantages and technical effects:

[0045] The multi-camera collaborative event detection method provided by this invention achieves seamless video acquisition and full-area monitoring of the entire highway segment through a standardized, cross-coverage deployment method, effectively overcoming the limitations of traditional systems with sparse deployment points and incomplete coverage. This method utilizes a multi-camera collaborative recognition and relay tracking mechanism to continuously and accurately locate and analyze the behavior of sudden abnormal targets such as spilled objects, pedestrians trespassing, and illegally parked vehicles, significantly improving the timeliness and accuracy of event detection. Simultaneously, its risk assessment model, which integrates multi-dimensional features, can intelligently quantify and classify the severity of events, providing scientific and intuitive decision support for command and dispatch and emergency response, thereby comprehensively improving the operational safety level and event response efficiency of highways. Attached Figure Description

[0046] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0047] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0048] Figure 2 This is a diagram illustrating the installation scenario of the device according to an embodiment of the present invention. Detailed Implementation

[0049] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0050] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0051] Example 1

[0052] like Figure 1 As shown, this embodiment provides a multi-camera collaborative event detection method for full coverage deployment on highways, including:

[0053] Multi-camera wide-area video acquisition devices are deployed intersecting along the target road segment to achieve full coverage video acquisition of the target road segment and obtain multiple video streams;

[0054] In the data analysis server, a camera processing unit is established for each camera to decode the video stream into image sequence frames, and each frame image, along with its acquisition timestamp and road segment number, is encapsulated into a raw data unit.

[0055] The image frames in the original data unit are subjected to ROI candidate region extraction to obtain the target candidate ROI region;

[0056] Target detection is performed on the candidate ROI region to obtain the target ROI coordinates, target type and confidence level, and target feature vector is extracted;

[0057] The target feature vector is input into the target tracking algorithm to update the target movement trajectory, and the target trajectory ID, target ROI coordinates, target type, confidence level, target feature vector and acquisition time are packaged into target detection data;

[0058] In each camera processing unit, local target tracking is performed based on the target detection data to establish a target tracking sequence;

[0059] The target tracking sequence is subjected to velocity trend analysis to obtain target velocity trend markers;

[0060] Based on the target velocity trend marker and target type, multi-camera collaborative tracking processing is performed;

[0061] Based on the collaborative tracking results, a risk assessment is performed on the target tracking sequence, and the risk level is output.

[0062] The workflow of this embodiment is mainly divided into six steps: full coverage deployment, data acquisition, target detection, target tracking, cross-camera collaboration, decision system evaluation and early warning.

[0063] Full coverage deployment: Multi-camera wide-area video acquisition devices can acquire video data within 1 kilometer by overlapping the field of view of adjacent cameras. By deploying them at 1-kilometer intervals along the side of the target highway, a full coverage network can be formed.

[0064] Raw data acquisition: The wide-area video acquisition device continuously acquires video data streams of the two-way lanes at a fixed frame rate through multiple built-in cameras, and encapsulates the video stream data with information such as the acquisition time of each frame, the acquisition section number, and the acquisition camera number into structured raw data units.

[0065] Target detection: For the video frame sequence and corresponding acquisition time of each frame in the raw data unit of a single camera, a background subtraction algorithm is used to obtain the relative coordinates of the foreground target in the image and separate the ROI region of the foreground target. For the target ROI image, a YOLOv10 neural network is used to obtain the target's category (discharged object, pedestrian, vehicle) and confidence level. A ConvNeXt-L neural network is used to extract the target's feature vector, and the DeepSORT algorithm is used to obtain the target's movement trajectory and state information. The YOLOv10 neural network needs to be trained using images acquired by the aforementioned wide-area video acquisition device and labeled with the target. The ConvNeXt-L neural network model is fine-tuned using cross-sectional images of the target to obtain cross-sectional matching capabilities. The DeepSORT algorithm directly uses the features extracted by the ConvNeXt-L model for target tracking.

[0066] Target tracking: Target trajectory ID, target ROI coordinates, target type, target confidence level, target feature vector, and target acquisition time are packaged into target detection data. Simultaneously, a target tracking sequence is established using the camera as the processing unit. Target tracking between cameras is achieved by combining target tracking ID and target feature vector similarity matching, and the target tracking sequence is maintained. Furthermore, target velocity trend analysis is performed on the target tracking sequence to obtain static, acceleration, and deceleration trend markers.

[0067] Cross-camera collaboration: In the event of sudden anomalies, a target tracking sequence is created, and the sequence continuously records event tracking segments generated by the event camera to achieve real-time tracking of the anomaly. Two types of cross-camera collaborative processing are performed based on target category confidence:

[0068] Calculate the spill risk correction parameter: When the abnormal event target is highly likely to be spilled material, the probability of the target moving is low. In addition to recording the event tracking segments from the event camera, the original target tracking sequence also incorporates the acceleration and deceleration trends of the target tracking sequences from all adjacent preceding cameras of the event camera to collaboratively confirm the spill risk coefficient. .

[0069] Cross-camera serial tracking: When an abnormal event target belongs to a pedestrian or vehicle and moves from the original camera's field of view to an overlapping area with the field of view of any adjacent camera, the camera uses feature matching to associate a newly created target tracking sequence with the adjacent camera, preventing duplicate event reporting. If the target leaves the overlapping area and returns to the original camera's field of view or enters the field of view of an adjacent camera, the original tracking sequence and the new target tracking sequence are spliced ​​together to achieve relay tracking of the moving target.

[0070] Risk assessment and early warning for weighted decision-making systems:

[0071] For the target tracking sequence, risk scoring is performed in real time for three events: pedestrian intrusion, illegal parking, and littering.

[0072] The pedestrian intrusion risk score X is calculated using the following formula: ,in This is the intrusion depth coefficient of the driving lane. The weights corresponding to the lane intrusion depth coefficient are: This represents the number of lane trajectory intersections. The weights corresponding to the lane trajectory intersection rate are... The time spent in the lane. The weights correspond to the lane dwell time.

[0073] The risk score Y for illegally parked vehicles is calculated using the following formula: ,in For lane encroachment, The weight corresponding to the shoulder encroachment degree. This is the parking duration coefficient. The weights corresponding to the parking duration coefficient are as follows: For the rate of change of vehicle speed, Weights corresponding to the rate of change of vehicle speed

[0074] The spill risk score Z is calculated using the following formula: ,in The relative size of the object. The weights correspond to the relative sizes of the objects. Duration of appearance Weights are assigned based on the duration of occurrence. For the sharpness of the outline, The weights corresponding to the sharpness of the contour.

[0075] Three different risk score thresholds were designed for pedestrian intrusion risk score X, illegally parked vehicle risk score Y, and spilled object risk score Z. The thresholds correspond to the first, second, and third level warning levels, respectively, to achieve alarms for different levels of severity.

[0076] The following is a detailed explanation using data:

[0077] Step 1: As Figure 2As shown, the multi-camera wide-area video acquisition device can achieve full coverage video acquisition of a target road segment of 1000m. The acquisition device includes a 12m high column and a camera group installed on top of the column. The video stream acquired by the camera group is required to be in H.264 format with a frame rate of 25fps. It includes one PTZ camera, two short-focus bullet cameras, and two long-focus bullet cameras. One PTZ camera acquires video over a 50m area on both sides of the column. The two short-focus bullet cameras are set back-to-back to acquire video over a 20m to 200m area on both sides of the column. The two long-focus bullet cameras are set back-to-back to acquire video over a 150m to 550m area on both sides of the column. To ensure full coverage of event detection, there must be overlap in the field of view between the multiple cameras in the camera group. There is a 30m overlap area between a single PTZ camera and an adjacent short-focus bullet camera, and a 50m overlap area between a short-focus bullet camera and an adjacent long-focus bullet camera. Multi-camera wide-area video acquisition devices need to be deployed crosswise along the target road segment, with a distance of less than 1000m between the acquisition devices to ensure that there is also an overlapping field of view between the devices, thereby ensuring full coverage of event detection.

[0078] Step 2: The data analysis server establishes a corresponding camera processing unit for each camera. Each camera processing unit decodes the H.264 video captured by that camera into RGB image sequence frames in real time. Then, each image frame, its acquisition timestamp (accurate to milliseconds), and the unique segment number Z of the acquisition device are encapsulated into a structured raw data unit.

[0079] Step 3: The camera processing unit extracts ROI candidate regions from the image: When a new frame of raw data is received, the corresponding image frame B is acquired. The background is continuously updated using the moving average background modeling method, and a background image D is acquired. Foreground difference operation is performed between the current image B and the background image D to obtain the difference image. Finally, the difference image is binarized and thresholded to extract connected components and obtain the target candidate ROI region.

[0080] Step 4: For the target candidate ROI region image extracted in Step 3, use the pre-trained YOLOV10 model to perform target detection, obtain the accurate target ROI coordinates, and parse the target type (pedestrian, vehicle, spilled object) and confidence level. Then, use the pre-trained ConvNeXt-L model to extract the target feature vector and feed it into the DeepSORT algorithm to continuously update the target movement trajectory.

[0081] The YOLOV10 neural network is trained using images collected by the aforementioned wide-area video acquisition devices and labeled with the target. The ConvNeXt-L neural network model is fine-tuned using cross-sectional images of the target to obtain cross-sectional matching capability.

[0082] Step 5: Package the target trajectory ID, target ROI coordinates, target type, target confidence score, target feature vector, and target acquisition time into target detection data. Multiple target detection data sets will be generated for each frame of the image.

[0083] Step 6: The camera processing unit performs target tracking for this camera. The steps are as follows:

[0084] Receive target detection data E;

[0085] If there is no target tracking sequence currently available, a target tracking sequence is created, and E is added to the target tracking sequence. The target tracking sequence is a collection of target detection data and maintains the tracking lifecycle of a target.

[0086] If a target tracking sequence already exists, iterate through all target tracking sequences. If the ID of the latest target detection data is the same as the trajectory ID in E, then add E to the target tracking sequence.

[0087] If E is not added to any target tracking sequence, iterate through each target tracking sequence, compare its features with E, and calculate the similarity. If the similarity is greater than the matching similarity threshold, then E is added to the target tracking sequence. The similarity calculation method is as follows: for the current target tracking sequence, calculate the cosine similarity between the target feature vectors in all target detection data and the target feature vectors of E, and take the average value, which is the similarity between the target tracking sequence and E.

[0088] If E has not yet been added to any target tracking sequence, it means that E is a new target. Create a target tracking sequence and add E to the target tracking sequence.

[0089] Step 7: Perform target velocity trend analysis on the target tracking sequence. The steps are as follows:

[0090] Using a statistical time period T (default is 5 seconds), extract the target detection data sets within the latest three T periods, and record them as sets according to their chronological order. , , (The target detection time span in each set is T,) The earliest, (Latest)

[0091] Calculate separately , , The pixel displacement of the ROI is obtained by using a pre-defined relationship between pixels and actual distances to determine the displacement distance of each set, and then dividing by time T to obtain the velocity. , , ,

[0092] when Less than the speed threshold If so, then mark the target as stationary;

[0093] Otherwise, when This indicates that the target is decelerating, and the target is marked as decelerating (where the coefficient 1.3 is configurable).

[0094] Otherwise, when This indicates that the target is accelerating, and the target is marked as accelerating (where the coefficient 1.3 is configurable).

[0095] Step 8: Multi-camera collaborative tracking processing. The steps are as follows:

[0096] Calculate the spill risk correction parameters:

[0097] For each camera processing unit, the following processing is performed: when the target in the target tracking sequence is a projectile, and the confidence level is ≥ the projectile threshold, and the target remains stationary for 2 consecutive seconds:

[0098] For the current camera and the previous camera at the actual installation point, determine the target speed trend. If either camera meets the following trend judgment condition, then mark the spillage risk correction parameter S of the previous target tracking sequence as the corresponding value in the table. The trend judgment condition is: within the most recent 1-minute statistical period, when the target type on the same lane is a vehicle, calculate the vehicle target speed trend (obtained from step 7). If the conditions in the table below are met, output the corresponding risk correction parameter. The risk correction parameters are shown in Table 1.

[0099] Table 1

[0100]

[0101] Cross-camera collaborative tracking:

[0102] When the pedestrian confidence score is greater than or equal to the pedestrian threshold, or the vehicle confidence score is greater than or equal to the vehicle threshold, and the target enters the overlapping area of ​​the field of view of adjacent cameras, the next adjacent camera at the actual installation point will also establish a new target tracking sequence for the same target. At this time, the feature similarity between the current target tracking sequence and the target tracking sequence of the previous camera is calculated. If the similarity is higher than the cross-camera matching threshold, the tracking sequences will be associated. The new target tracking sequence will not be used for subsequent event decision evaluation to prevent repeated event reporting. The feature similarity calculation method is as follows: for each target feature vector in one target tracking sequence, the cosine similarity is calculated with each target feature vector in another target tracking sequence, and the mean value is calculated.

[0103] When a target leaves the overlapping area of ​​the next camera's field of view and returns to the field of view of the previous camera, the new target tracking sequence is discarded, and the existing target tracking sequence continues to be used to track the target. When a target leaves the overlapping area of ​​the adjacent camera's field of view and moves to the field of view of the next adjacent camera, the redundant target tracking data generated by the existing tracking sequence in the overlapping area of ​​the field of view is discarded. At the same time, subsequent event decision evaluation can be performed, and the target tracking sequences recorded by adjacent cameras and the existing target tracking sequences are fused and stitched together according to the time sequence to achieve relay tracking of dynamic targets.

[0104] Step 9: Risk assessment and early warning for the weighted decision-making system.

[0105] Each camera processing unit iterates through all target tracking sequences and takes the highest analysis period threshold. The target detection dataset K (within one minute by default) is used to output a total risk score through a multi-dimensional weighted fusion algorithm. Different levels of early warning are achieved by setting different risk score thresholds. The main process is as follows:

[0106] If the most recent target detection data type in the target tracking sequence is a pedestrian, and the confidence level is greater than or equal to the pedestrian threshold, calculate the pedestrian intrusion risk score based on the pedestrian intrusion risk behavior characteristics. The weighting is shown in Table 2:

[0107] Table 2

[0108]

[0109] The lane intrusion depth coefficient R is calculated as follows:

[0110] If the pedestrian is already in the driving lane (not the emergency lane or emergency stopping area), then record R=1;

[0111] Otherwise, calculate the nearest lane pixel width Rw, and the horizontal pixel distance d from the bottom center point of the target ROI rectangle to the driving lane, then R = d / Rw.

[0112] The number of lane trajectory intersections, Sc, is the number of pedestrians as the current target type.

[0113] The lane dwell time ratio Rt is calculated as follows: In the target detection dataset K, find the time difference dt between the first moment and the current moment for targets of the pedestrian type, and let...

[0114] A risk score for pedestrian intruders is calculated using a weighted average. .

[0115] If the most recent target detection data type in the target tracking sequence is a vehicle, and the target speed trend is stationary,

[0116] Calculate the risk score for illegally parked vehicles based on their risk behavior characteristics. The weight allocation is shown in Table 3:

[0117] Table 3

[0118]

[0119] The lane encroachment degree O is calculated as follows: determine the lane position of the bottom center point of the stationary vehicle's ROI. If it is in the driving lane, O is set to 1; otherwise, it is set to 0.

[0120] The parking duration coefficient P is calculated as follows: In the target detection data set K of illegally parked vehicles, find the time difference dt from the first moment when the target speed trend is stationary to the current moment, and let...

[0121] The rate of change of vehicle speed Q is calculated as follows: starting from the vertical coordinate of the ROI position of illegally parked vehicles, and extending to the lane area opposite to the direction of vehicle travel, take the total number of target vehicles as Cc, and their target speed trend as the number of decelerations as Cd. Then Q = Cc ÷ Cd.

[0122] A weighted calculation is used to determine the risk score for illegal parking. .

[0123] If the most recent target detection data type in the target tracking sequence is projectile, and the confidence level is greater than or equal to the projectile threshold, calculate the projectile risk score based on the projectile behavior characteristics. The weight allocation is shown in Table 4:

[0124] Table 4

[0125]

[0126] The relative size U of an object is calculated as follows:

[0127] If the spilled material is in the emergency lane or emergency parking area, then record U=0.1;

[0128] Otherwise, calculate the nearest lane pixel width Rw, and the target ROI size s = max(rw, rh), where rw is the ROI width and rh is the ROI height. Calculate U = s / Rw.

[0129] The duration V of occurrence is calculated as follows: In the target detection data set K of illegally parked vehicles, find the time difference dt from the first moment when the target speed trend is stationary to the current moment, and let... .

[0130] Contour sharpness W, calculation method:

[0131] Canny edge detection is performed on the ROI region image, and edge detection is performed on the generated Canny edge map to obtain the edge pixel set of the target region. Based on the edge pixel set, closed boundaries in the image are extracted to obtain a contour curve representing the target shape. The contour curve is then simplified; specifically, based on a preset approximation accuracy, polygon fitting is performed on the point set of the contour curve to obtain a polygonal approximate contour composed of several vertices. A vertex set Pt is obtained, with a number of vertices. For each point coordinate in the vertex set Pt, take an image region with a width and height of 20 centered at that coordinate. Calculate the corner points using the Shi-Tomasi method. If the corner point coordinate is the center, then that vertex is a sharp vertex. Let the number of sharp vertices be denoted as . , through formula Calculate the profile sharpness W.

[0132] The risk score of spilled material is obtained through weighted calculation. .in The spill risk correction parameters are obtained in step 8.

[0133] Risk classification and early warning

[0134] Three levels of early warning are set for pedestrian risk, vehicle illegal parking risk, and debris spillage risk, along with corresponding risk warning thresholds.

[0135] Level 1 pedestrian risk warning is the highest level, and it is triggered when the pedestrian risk score is... Triggered at time The default value is 0.8;

[0136] The pedestrian risk warning level is level two, which is medium. (When the pedestrian risk score...) Triggered at time The default value is 0.6;

[0137] The pedestrian risk warning level is three, which is low. (When the pedestrian risk score is...) Triggered at time The default value is 0.3;

[0138] Level 1 Warning for Illegally Parked Vehicles, the highest level, is issued when the risk score for illegally parked vehicles is [not specified]. Triggered at time The default value is 1.0;

[0139] Level 2 Warning for Illegally Parked Vehicles: This is a medium-level warning. (The text abruptly ends here, likely due to an incomplete sentence or missing information.) Triggered at time The default value is 0.8;

[0140] The risk level for illegally parked vehicles is three-tiered, which is considered low. (The risk score for illegally parked vehicles is...) Triggered at time The default value is 0.6;

[0141] Level 1 warning for spillage risk is the highest level, when the spillage risk score is... Triggered at time The default value is 1.0;

[0142] Level II alert for spillage risk, classified as medium, when the spillage risk score... Triggered at time The default value is 0.7;

[0143] The spill risk warning is at level three, which is low. (When the spill risk score is...) Triggered at time The default value is 0.5;

[0144] Personnel can implement different response plans based on the alarm level.

[0145] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-camera collaborative event detection method for full coverage deployment on highways, characterized in that, include: Multi-camera wide-area video acquisition devices are deployed intersecting along the target road segment to achieve full coverage video acquisition of the target road segment and obtain multiple video streams; In the data analysis server, a camera processing unit is established for each camera to decode the video stream into image sequence frames, and each frame image, along with its acquisition timestamp and road segment number, is encapsulated into a raw data unit. The image frames in the original data unit are subjected to ROI candidate region extraction to obtain the target candidate ROI region; Target detection is performed on the candidate ROI region to obtain the target ROI coordinates, target type and confidence level, and target feature vector is extracted; The target feature vector is input into the target tracking algorithm to update the target movement trajectory, and the target trajectory ID, target ROI coordinates, target type, confidence level, target feature vector and acquisition time are packaged into target detection data; In each camera processing unit, local target tracking is performed based on the target detection data to establish a target tracking sequence; The target tracking sequence is subjected to velocity trend analysis to obtain target velocity trend markers; Based on the target velocity trend marker and target type, multi-camera collaborative tracking processing is performed; Multi-camera cooperative tracking processing includes: Based on the target type and confidence level, determine whether it is spilled material, a pedestrian, or a vehicle; Based on the target velocity trend markers and the target tracking sequences of adjacent cameras, the projectile risk correction parameters are calculated; Based on the situation where the target enters the overlapping area of ​​the field of view of adjacent cameras, cross-camera target tracking sequences are associated and stitched together; Based on the collaborative tracking results, a risk assessment is performed on the target tracking sequence, and the risk level is output.

2. The method according to claim 1, characterized in that, The multi-camera wide-area video acquisition device includes a column and a camera group installed on the top of the column. The camera group includes a PTZ camera, a short-focus bullet camera, and a long-focus bullet camera. Based on the deployment positions of the PTZ camera, short-focus camera, and long-focus camera, video data for different areas is acquired; Based on the overlapping areas of the fields of view between adjacent cameras, obtain the cross-over video data.

3. The method according to claim 1, characterized in that, Extracting ROI candidate regions from the image frames in the original data unit includes: Update the background image using the moving average background modeling method; Obtain the difference image based on the difference operation between the current image frame and the background image; Based on the binarization thresholding of the difference image, connected components are extracted as candidate ROI regions.

4. The method according to claim 1, characterized in that, Target detection of the candidate ROI region includes: Based on the pre-trained YOLOV10 model, the target candidate ROI region is detected, and the target ROI coordinates, target type and confidence score are obtained; Based on the pre-trained ConvNeXt-L model, feature extraction is performed on the target ROI region to obtain the target feature vector.

5. The method according to claim 1, characterized in that, Inputting the target feature vector into the target tracking algorithm includes: Based on the DeepSORT algorithm, the target's trajectory is updated using the target's feature vector. Target detection data is generated based on the target trajectory ID, target ROI coordinates, target type, confidence level, target feature vector, and acquisition time.

6. The method according to claim 1, characterized in that, Local target tracking in each camera processing unit includes: Based on the target trajectory ID in the target detection data, add the corresponding data to the existing target tracking sequence; Based on the similarity between target feature vectors, unmatched target detection data are matched with existing target tracking sequences; Update or create a new target tracking sequence based on the matching results.

7. The method according to claim 1, characterized in that, Velocity trend analysis of target tracking sequences includes: Extract multiple target detection datasets based on time periods; Based on the calibration relationship between pixels and actual distance, calculate the displacement distance and velocity of the target in each cycle; Based on the speed change, determine whether the target is stationary, accelerating, or decelerating, and generate a speed trend marker.

8. The method according to claim 1, characterized in that, Risk assessment of target tracking sequences includes: Based on the target type, calculate the risk score for pedestrians who break in, the risk score for illegally parked vehicles, and the risk score for debris; Based on the comparison between the risk score and the preset threshold, the corresponding risk level warning is triggered.

Images