Law enforcement video intelligent analysis method based on deep learning
By employing deep learning methods for rapid edge detection and fine-grained analysis, combined with adaptive enhancement processing and behavior tracking technology, this approach addresses the issues of low detection accuracy and real-time performance in complex scenarios in traditional law enforcement video analysis, achieving highly efficient law enforcement video analysis.
Patent Information
- Application Number
- CN202511317888.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-16
- Publication Date
- 2025-12-16
AI Technical Summary
Traditional law enforcement video analysis methods have low detection accuracy in complex scenarios, lack customized processing, and the edge computing and cloud collaboration architecture has not been fully implemented, resulting in high data transmission latency and failing to meet the needs of real-time law enforcement.
A deep learning-based intelligent analysis method for law enforcement videos is adopted. It quickly identifies anomaly types at the edge and uploads them to the cloud for fine-grained detection. Combined with adaptive enhancement processing and behavior tracking technology, it dynamically adjusts the local features of video frames to achieve stable cross-frame tracking and multi-dimensional feature analysis.
It reduces data transmission volume, shortens anomaly response time to the second level, improves the accuracy of single-target and group anomaly identification, meets the real-time requirements of law enforcement scenarios, and enhances the accuracy of target detection in complex scenarios.
Smart Images
Figure CN121147818A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and more specifically to a deep learning-based intelligent analysis method for law enforcement videos. Background Technology
[0002] With the deep integration of artificial intelligence and video surveillance technology, law enforcement video analytics, as a core application in fields such as intelligent transportation and public safety, directly impacts the timeliness and accuracy of law enforcement decisions in terms of efficiency and precision. Traditional law enforcement video analytics primarily relies on manual review, requiring staff to examine surveillance footage frame by frame to identify abnormal events and determine responsible parties.
[0003] With the increasing coverage of surveillance cameras, the amount of law enforcement video data has exploded, and traditional methods can no longer meet the law enforcement needs of "real-time response and accurate judgment." Although existing computer vision-based automated analysis technologies have been applied, they have obvious limitations: First, target detection and tracking algorithms are not adaptable enough to complex scenes. For example, in low-light environments, the contrast between the target and the background decreases, leading to a sharp drop in detection accuracy. Second, there is a lack of customized processing procedures for law enforcement scenarios, and general video analysis models struggle to accurately identify specific anomaly types. Third, the abnormal behavior and responsibility determination processes rely heavily on fixed rules, failing to incorporate dynamic information such as target movement trajectories and interaction characteristics, resulting in insufficient objectivity and persuasiveness in the judgment results.
[0004] Furthermore, the architecture of edge computing and cloud collaboration has not yet been fully implemented, with most systems relying solely on cloud processing, resulting in high data transmission latency and failing to meet real-time law enforcement needs. Simultaneously, video enhancement processing often employs global parameter adjustments, failing to adaptively optimize based on local intra-frame features, thus impacting the accuracy of subsequent target detection. These issues collectively restrict the level of intelligence in law enforcement video analysis. Summary of the Invention
[0005] To overcome the aforementioned deficiencies in the prior art, this invention provides a deep learning-based intelligent analysis method for law enforcement videos to address the problems existing in the background art.
[0006] This invention provides the following technical solution: a deep learning-based intelligent analysis method for law enforcement videos, comprising the following steps:
[0007] Step S01: Divide the law enforcement video into time periods and obtain videos of abnormal time periods: Obtain law enforcement videos of the traffic monitoring coverage area through surveillance cameras, and divide the videos into videos of normal time periods and videos of abnormal time periods;
[0008] Step S02: Extract the video of the time period to be analyzed and perform adaptive enhancement processing on the video frames: establish a video set of abnormal time periods, wherein the abnormal time period videos in the video set are the videos of the time period to be analyzed, and perform adaptive enhancement processing on the video frames of the time period to be analyzed.
[0009] Step S03: Combine edge processing technology to perform target detection on the enhanced video frame: Combine the edge processing technology to quickly detect and identify the anomaly type of the video frame, and upload the anomaly type identified by the edge processing to the cloud for fine-grained detection to obtain the number of target objects;
[0010] Step S04: Matching adaptive behavior tracking technology based on target detection results: Matching adaptive behavior tracking technology according to the number of target objects, wherein the adaptive behavior tracking technology includes single target behavior tracking and multiple target behavior tracking;
[0011] Step S05: Comprehensive analysis of adaptive behavior tracking results to determine responsibility for abnormal behavior: Comprehensive analysis of single target behavior tracking results and multiple target behavior tracking results is performed to complete the intelligent analysis of law enforcement videos and output the result of determining responsibility for abnormal behavior.
[0012] Preferably, in step S01, the specific content of dividing the law enforcement video into time periods and obtaining videos from abnormal time periods is as follows:
[0013] Enforcement videos are acquired by using surveillance cameras to monitor traffic coverage areas. The acquired enforcement videos include the camera number and time period corresponding to the video. The surveillance cameras are connected to the traffic management center's server and transmit video data in real time via the network.
[0014] The acquired law enforcement videos are divided into regular time period videos and abnormal time period videos according to preset division rules. After verifying their identity, the management personnel can retrieve the surveillance videos of the specific time period.
[0015] Preferably, in step S02, the specific content of extracting the video frames for the time period to be analyzed and performing adaptive enhancement processing is as follows:
[0016] Establish a video set for abnormal time periods, wherein the videos in the video set for abnormal time periods are the videos for the time periods to be analyzed;
[0017] The contrast of video frames in the video set of abnormal time periods is adaptively adjusted by a local contrast enhancement algorithm.
[0018] Adaptive noise reduction is performed on video frames of the time period to be analyzed in the video set of abnormal time periods;
[0019] Texture detection is performed on video frames within the video set for the period to be analyzed in the abnormal time period. If the texture is blurry, the edges are enhanced in a targeted manner using a sharpening algorithm.
[0020] Preferably, in step S03, the specific content of target detection in the enhanced video frame combined with edge processing technology is as follows:
[0021] Edge contours in video frames are extracted using an edge detection algorithm, and edge gradient directions are obtained to filter out valid edge regions that are not part of the background in the video frames.
[0022] The effective edge region is matched with the preset anomaly type feature library. The anomaly type is determined by calculating the feature similarity. If the calculated feature similarity is greater than or equal to the preset matching threshold, the anomaly type is successfully identified. Otherwise, if the calculated feature similarity is less than the preset matching threshold, the anomaly type is not successfully identified, and the system returns to the anomaly type feature library for rematching.
[0023] The edge-end outputs an anomaly type label based on the anomaly type determination result, and marks the bounding box coordinates of the anomaly region;
[0024] The edge device uploads the anomaly type label, the coordinates of the anomaly region bounding box, and the corresponding video frame to the cloud. The cloud then calls a high-precision detection model to perform fine-grained detection and outputs the number of target objects, the coordinates of each target object, and the confidence level.
[0025] Preferably, the cloud-based high-precision detection model is invoked for fine-grained detection, and the specific content of the output, including the number of target objects, the coordinates of each target object, and the confidence level, is as follows:
[0026] Based on the bounding box coordinates of the marked abnormal regions, coarse identification of target objects is performed in the marked abnormal regions, and the number n of candidate target objects in the abnormal regions is obtained using lightweight target detection technology;
[0027] Extract the coordinates of each candidate target, which are represented in the form of a bounding box, including the coordinates of the top left corner and the bottom right corner;
[0028] Based on the bounding box analysis of each candidate target, the degree of overlap between the bounding box of the candidate target and the predefined template box is analyzed, and the confidence score of each candidate target is calculated.
[0029] A high-precision detection model is invoked to perform fine-grained detection on each candidate target and output the number of target objects. The high-precision detection model consists of indicator functions.
[0030] Preferably, in step S04, the specific content of the adaptive behavior tracking technology based on the target detection result is as follows:
[0031] The tracking strategy is dynamically switched based on the number of candidate target objects m output: when m=1, single target behavior tracking is enabled, and when m≥1, multiple target behavior tracking is enabled.
[0032] When single target behavior tracking is enabled, the target's position coordinates in each frame of video are recorded in coordinate form to form a single trajectory sequence. The target's speed, direction of movement, and stationary state are also recorded, and a single target behavior label is automatically generated.
[0033] When multiple target behavior tracking is enabled, a unique tracking ID is assigned to each target. Based on the unique ID of each target, the position coordinates of the target in each frame of video are obtained to form multiple trajectory sequences. The distance changes, relative motion directions and interaction frequencies between multiple targets are recorded, and multiple target interaction tags are automatically generated.
[0034] Preferably, in step S05, the specific content of determining responsibility for abnormal behavior based on the comprehensive analysis of the adaptive behavior tracking results is as follows:
[0035] Based on the tracking of individual target behavior, a single target behavior label is automatically generated. The target's speed, direction of movement, and dwell state are comprehensively analyzed. When the target's speed, direction of movement, or dwell state is outside the preset threshold range, it is marked as a potential abnormal behavior. When a single target is marked as abnormal, the single target corresponding to the trajectory sequence is identified as the subject of prediction responsibility.
[0036] Based on the tracking of multiple target behaviors, multiple target interaction tags are automatically generated. The distance changes, relative motion directions and interaction frequencies between multiple targets are comprehensively analyzed. When the distance changes, relative motion directions or interaction frequencies between multiple targets exceed the preset reasonable range, they are marked as potential abnormal group behavior. When multiple targets are marked as abnormal, the multiple targets with the highest interaction frequency in the trajectory sequence are identified as the predicted responsible parties.
[0037] The technical effects and advantages of this invention are as follows:
[0038] This invention rapidly identifies anomaly types through edge detection, focuses on fine-grained analysis and target counting, and forms an efficient collaborative mode of edge coarse screening and cloud fine calculation. This architecture not only reduces the amount of data transmission, but also shortens the anomaly response time to the second level, meeting the real-time requirements of law enforcement scenarios.
[0039] Based on a dynamic switching tracking strategy for the number of targets, combined with deep learning features, stable cross-frame tracking is achieved. At the same time, by analyzing abnormal behavior through multi-dimensional features such as trajectory sequence and interaction frequency, the accuracy of single-target anomaly identification and group anomaly identification is greatly improved.
[0040] By employing adaptive enhancement processing technology, the enhancement strategy is dynamically adjusted based on the local features of video frames, effectively addressing issues such as low light, noise, and occlusion. While preserving target details, noise is suppressed, thus improving target detection accuracy in complex scenes. Attached Figure Description
[0041] Figure 1 This is a flowchart illustrating a deep learning-based intelligent analysis method for law enforcement videos. Detailed Implementation
[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The deep learning-based intelligent analysis method for law enforcement videos involved in the present invention is not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] like Figure 1 As shown, this invention provides a deep learning-based intelligent analysis method for law enforcement videos, comprising the following steps:
[0044] Step S01: Divide the law enforcement video into time periods and obtain videos of abnormal time periods: Obtain law enforcement videos of the traffic monitoring coverage area through surveillance cameras, and divide the videos into videos of normal time periods and videos of abnormal time periods;
[0045] Step S02: Extract the video of the time period to be analyzed and perform adaptive enhancement processing on the video frames: establish a video set of abnormal time periods, wherein the abnormal time period videos in the video set are the videos of the time period to be analyzed, and perform adaptive enhancement processing on the video frames of the time period to be analyzed.
[0046] Step S03: Combine edge processing technology to perform target detection on the enhanced video frame: Combine the edge processing technology to quickly detect and identify the anomaly type of the video frame, and upload the anomaly type identified by the edge processing to the cloud for fine-grained detection to obtain the number of target objects;
[0047] Step S04: Matching adaptive behavior tracking technology based on target detection results: Matching adaptive behavior tracking technology according to the number of target objects, wherein the adaptive behavior tracking technology includes single target behavior tracking and multiple target behavior tracking;
[0048] Step S05: Comprehensive analysis of adaptive behavior tracking results to determine responsibility for abnormal behavior: Comprehensive analysis of single target behavior tracking results and multiple target behavior tracking results is performed to complete the intelligent analysis of law enforcement videos and output the result of determining responsibility for abnormal behavior.
[0049] In this embodiment, it should be specifically noted that in step S01, the specific content of dividing the law enforcement video into time periods and obtaining videos from abnormal time periods is as follows:
[0050] Enforcement videos are acquired by using surveillance cameras to monitor traffic coverage areas. The acquired enforcement videos include the camera number and time period corresponding to the video. The surveillance cameras are connected to the traffic management center's server and transmit video data in real time via the network.
[0051] The acquired law enforcement videos are divided into regular time period videos and abnormal time period videos according to preset division rules. The preset division rules refer to specific time period division standards set based on historical data and real-time conditions. For example, when a traffic accident occurs, the time period when the traffic flow is significantly lower or higher than the average level is classified as an abnormal time period video, and the rest is classified as a regular time period video. In this embodiment, when the deviation between the real-time traffic flow and the historical average traffic flow for the same period exceeds 30%, it is marked as an abnormal time period. When events such as traffic accidents, illegal parking, and pedestrian intrusion are detected, the 30 minutes before and after the event are automatically classified as abnormal time periods. After verifying their identity, the management personnel can retrieve the monitoring video of the specific time period.
[0052] In this embodiment, it should be specifically explained that the specific content of step S02, which involves extracting the video frames of the time period to be analyzed and performing adaptive enhancement processing, is as follows:
[0053] Establish a video set for abnormal time periods, wherein the videos in the video set for abnormal time periods are the videos for the time periods to be analyzed;
[0054] The contrast of video frames in the video set of abnormal time periods is adaptively adjusted by a local contrast enhancement algorithm, and the details in dark or bright areas are magnified by the local contrast enhancement algorithm.
[0055] Adaptive noise reduction is performed on video frames of the time period to be analyzed in the video set of abnormal time periods;
[0056] Texture detection is performed on video frames of the time period to be analyzed in the video set of abnormal time periods. If the texture is blurry, the edges are enhanced in a targeted manner through a sharpening algorithm, while the sharpening of the background is suppressed to prevent cluttered background information from standing out.
[0057] In this embodiment, it should be specifically explained that the specific content of the target detection of the enhanced video frame in step S03, which combines edge processing technology, is as follows:
[0058] Edge contours in video frames are extracted using an edge detection algorithm, and edge gradient directions are obtained to filter out valid edge regions that are not part of the background in the video frames.
[0059] The effective edge region is matched with the preset anomaly type feature library. The anomaly type is determined by calculating the feature similarity. If the calculated feature similarity is greater than or equal to the preset matching threshold, the anomaly type is successfully identified. Otherwise, if the calculated feature similarity is less than the preset matching threshold, the anomaly type is not successfully identified, and the system returns to the anomaly type feature library for rematching.
[0060] The edge-end outputs an anomaly type label based on the anomaly type determination result, and marks the bounding box coordinates of the anomaly region;
[0061] The edge device uploads the anomaly type label, the coordinates of the anomaly region bounding box, and the corresponding video frame to the cloud. The cloud then calls a high-precision detection model to perform fine-grained detection and outputs the number of target objects, the coordinates of each target object, and the confidence level.
[0062] In this embodiment, it should be specifically noted that the cloud-based high-precision detection model performs fine-grained detection, and the specific content of the output, including the number of target objects, the coordinates of each target object, and the confidence level, is as follows:
[0063] Based on the bounding box coordinates of the marked abnormal regions, coarse identification of target objects is performed in the marked abnormal regions, and the number n of candidate target objects in the abnormal regions is obtained using lightweight target detection technology;
[0064] Extract the coordinates of each candidate target, which are represented in the form of a bounding box, including the coordinates of the top left corner and the bottom right corner;
[0065] Based on the bounding box analysis of each candidate target, the overlap between the candidate target's bounding box and the predefined template box is calculated, and the confidence score of each candidate target is calculated using the following formula: c i =IoU(B i G i )×P i , where c i B represents the confidence level of each candidate target. i G represents the bounding box of each candidate target. i P represents the predefined template box corresponding to the bounding box of each candidate target. i This represents the predicted probability of each candidate target belonging to a certain category, such as the probability of a person or a car. The value range is [0,1], where i = 1, 2, 3, ..., n, and i represents the candidate target number in the anomaly region; IoU(B i G i The intersection-union ratio (IU) represents the degree of overlap between two bounding boxes. The formula for IU is: The value range is [0,1];
[0066] A high-precision detection model is invoked to perform fine-grained detection on each candidate target, and the number of target objects is output. The high-precision detection model consists of indicator functions, and the calculation formula is as follows: Where m represents the number of target objects in the fine-grained detection output, n represents the number of candidate target objects in the abnormal region, and I(c i ≥θ) represents the indicator function. When the confidence level of the candidate target is greater than or equal to the preset threshold, the indicator function outputs 1, and otherwise the indicator function outputs 0.
[0067] In this embodiment, it should be specifically explained that the specific content of the adaptive behavior tracking technology based on the target detection result in step S04 is as follows:
[0068] The tracking strategy is dynamically switched based on the number of candidate target objects m output: when m=1, single target behavior tracking is enabled, and when m≥1, multiple target behavior tracking is enabled.
[0069] When single target behavior tracking is enabled, the target's position coordinates in each frame of video are recorded in coordinate form to form a single trajectory sequence. The target's speed, direction of movement, and stationary state are also recorded, and a single target behavior label is automatically generated.
[0070] When multiple target behavior tracking is enabled, a unique tracking ID is assigned to each target. Based on the unique ID of each target, the position coordinates of the target in each frame of video are obtained to form multiple trajectory sequences. The distance changes, relative motion directions and interaction frequencies between multiple targets are recorded, and multiple target interaction tags are automatically generated.
[0071] In this embodiment, it should be specifically explained that the specific content of determining the responsibility for abnormal behavior based on the comprehensive analysis of the adaptive behavior tracking results in step S05 is as follows:
[0072] Based on the tracking of individual target behavior, a single target behavior label is automatically generated. The target's speed, direction of movement, and dwell state are comprehensively analyzed. When the target's speed, direction of movement, or dwell state is outside the preset threshold range, it is marked as a potential abnormal behavior. When a single target is marked as abnormal, the single target corresponding to the trajectory sequence is identified as the subject of prediction responsibility.
[0073] Based on the tracking of multiple target behaviors, multiple target interaction tags are automatically generated. The distance changes, relative motion directions and interaction frequencies between multiple targets are comprehensively analyzed. When the distance changes, relative motion directions or interaction frequencies between multiple targets exceed the preset reasonable range, they are marked as potential abnormal group behavior. When multiple targets are marked as abnormal, the multiple targets with the highest interaction frequency in the trajectory sequence are identified as the predicted responsible parties.
[0074] The main difference between this implementation and the existing technology is that this embodiment quickly identifies anomaly types through edge detection, focuses on fine-grained analysis and target counting, and forms an efficient collaborative mode of edge coarse screening and cloud fine calculation. This architecture not only reduces the amount of data transmission, but also shortens the anomaly response time to the second level, meeting the real-time requirements of law enforcement scenarios.
[0075] Based on a dynamic switching tracking strategy for the number of targets, combined with deep learning features, stable cross-frame tracking is achieved. At the same time, by analyzing abnormal behavior through multi-dimensional features such as trajectory sequence and interaction frequency, the accuracy of single-target anomaly identification and group anomaly identification is greatly improved.
[0076] By employing adaptive enhancement processing technology, the enhancement strategy is dynamically adjusted based on the local features of video frames, effectively addressing issues such as low light, noise, and occlusion. While preserving target details, noise is suppressed, thus improving target detection accuracy in complex scenes.
[0077] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
[0078] The above description is merely a specific embodiment 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 deep learning-based intelligent analysis method for law enforcement videos, characterized in that: Includes the following steps: Step S01: Divide the law enforcement video into time periods and obtain videos of abnormal time periods: Obtain law enforcement videos of the traffic monitoring coverage area through surveillance cameras, and divide the videos into videos of normal time periods and videos of abnormal time periods; Step S02: Extract the video of the time period to be analyzed and perform adaptive enhancement processing on the video frames: establish a video set of abnormal time periods, wherein the abnormal time period videos in the video set are the videos of the time period to be analyzed, and perform adaptive enhancement processing on the video frames of the time period to be analyzed. Step S03: Combine edge processing technology to perform target detection on the enhanced video frame: Combine the edge processing technology to quickly detect and identify the anomaly type of the video frame, and upload the anomaly type identified by the edge processing to the cloud for fine-grained detection to obtain the number of target objects; Step S04: Matching adaptive behavior tracking technology based on target detection results: Matching adaptive behavior tracking technology according to the number of target objects, wherein the adaptive behavior tracking technology includes single target behavior tracking and multiple target behavior tracking; Step S05: Comprehensive analysis of adaptive behavior tracking results to determine responsibility for abnormal behavior: Comprehensive analysis of single target behavior tracking results and multiple target behavior tracking results is performed to complete the intelligent analysis of law enforcement videos and output the result of determining responsibility for abnormal behavior.
2. The method for intelligent analysis of law enforcement videos based on deep learning according to claim 1, characterized in that: In step S01, the law enforcement video is divided into time periods, and the specific content of videos from abnormal time periods is obtained as follows: Enforcement videos are acquired by using surveillance cameras to monitor traffic coverage areas. The acquired enforcement videos include the camera number and time period corresponding to the video. The surveillance cameras are connected to the traffic management center's server and transmit video data in real time via the network. The acquired law enforcement videos are divided into regular time period videos and abnormal time period videos according to preset division rules. After verifying their identity, the management personnel can retrieve the surveillance videos of the specific time period.
3. The method for intelligent analysis of law enforcement videos based on deep learning according to claim 1, characterized in that: In step S02, the specific details of extracting the video frames for the time period to be analyzed and performing adaptive enhancement processing are as follows: Establish a video set for abnormal time periods, wherein the videos in the video set for abnormal time periods are the videos for the time periods to be analyzed; The contrast of video frames in the video set of abnormal time periods is adaptively adjusted by a local contrast enhancement algorithm. Adaptive noise reduction is performed on video frames of the time period to be analyzed in the video set of abnormal time periods; Texture detection is performed on video frames within the video set for the period to be analyzed in the abnormal time period. If the texture is blurry, the edges are enhanced in a targeted manner using a sharpening algorithm.
4. The method for intelligent analysis of law enforcement videos based on deep learning according to claim 1, characterized in that: In step S03, the specific details of performing target detection on the enhanced video frame using edge processing technology are as follows: Edge contours in video frames are extracted using an edge detection algorithm, and edge gradient directions are obtained to filter out valid edge regions that are not part of the background in the video frames. The effective edge region is matched with the preset anomaly type feature library. The anomaly type is determined by calculating the feature similarity. If the calculated feature similarity is greater than or equal to the preset matching threshold, the anomaly type is successfully identified. Otherwise, if the calculated feature similarity is less than the preset matching threshold, the anomaly type is not successfully identified, and the system returns to the anomaly type feature library for rematching. The edge-end outputs an anomaly type label based on the anomaly type determination result, and marks the bounding box coordinates of the anomaly region; The edge device uploads the anomaly type label, the coordinates of the anomaly region bounding box, and the corresponding video frame to the cloud. The cloud then calls a high-precision detection model to perform fine-grained detection and outputs the number of target objects, the coordinates of each target object, and the confidence level.
5. The method for intelligent analysis of law enforcement videos based on deep learning according to claim 4, characterized in that: The cloud-based high-precision detection model performs fine-grained detection, and outputs the number of target objects, the coordinates of each target object, and the confidence level as follows: Based on the bounding box coordinates of the marked abnormal regions, coarse identification of target objects is performed in the marked abnormal regions, and the number n of candidate target objects in the abnormal regions is obtained using lightweight target detection technology; Extract the coordinates of each candidate target, which are represented in the form of a bounding box, including the coordinates of the top left corner and the bottom right corner; Based on the bounding box analysis of each candidate target, the degree of overlap between the bounding box of the candidate target and the predefined template box is analyzed, and the confidence score of each candidate target is calculated. A high-precision detection model is invoked to perform fine-grained detection on each candidate target and output the number of target objects. The high-precision detection model consists of indicator functions.
6. The method for intelligent analysis of law enforcement videos based on deep learning according to claim 1, characterized in that: In step S04, the specific content of the adaptive behavior tracking technology based on the target detection results is as follows: The tracking strategy is dynamically switched based on the number of candidate target objects m output: when m=1, single target behavior tracking is enabled, and when m≥1, multiple target behavior tracking is enabled. When single target behavior tracking is enabled, the target's position coordinates in each frame of video are recorded in coordinate form to form a single trajectory sequence. The target's speed, direction of movement, and stationary state are also recorded, and a single target behavior label is automatically generated. When multiple target behavior tracking is enabled, a unique tracking ID is assigned to each target. Based on the unique ID of each target, the position coordinates of the target in each frame of video are obtained to form multiple trajectory sequences. The distance changes, relative motion directions and interaction frequencies between multiple targets are recorded, and multiple target interaction tags are automatically generated.
7. The method for intelligent analysis of law enforcement videos based on deep learning according to claim 1, characterized in that: In step S05, the specific content of determining responsibility for abnormal behavior based on the comprehensive analysis of adaptive behavior tracking results is as follows: Based on the tracking of individual target behavior, a single target behavior label is automatically generated. The target's movement speed, movement direction, and dwell state are comprehensively analyzed. When the target's movement speed, movement direction, or dwell state is outside the preset threshold range, it is marked as a potential abnormal behavior. When a single target is marked as abnormal, the single target corresponding to the trajectory sequence is identified as the subject of prediction responsibility. Based on the tracking of multiple target behaviors, multiple target interaction tags are automatically generated. The distance changes, relative motion directions, and interaction frequencies between multiple targets are comprehensively analyzed. When the distance changes, relative motion directions, or interaction frequencies between multiple targets all exceed the preset reasonable range, they are marked as potential abnormal group behavior. When multiple targets are labeled as anomalies, the targets with the highest interaction frequency in the trajectory sequence are identified as the subjects responsible for prediction.