A police supervision abnormal behavior early warning method based on behavior feature recognition

Abstract

The application discloses a kind of based on behavior characteristic identification's police supervision abnormal behavior early warning method, comprising: S1, acquisition multi-modal flow data and reconstruct to visual flow anti-shielding skeleton topology;S2, compliance order is graphed and mapped as compliance boundary hyperplane;S3, joint multi-modal feature calculates rule intention evolution transfer probability, and output pre-alarm signal;S4, constructs dynamic context state machine driven hyperplane self-adapting drift, and output dynamic determination threshold;S5, check early warning and introduce the Grad-CAM backward decoding attribution of improved, extract Top-K key feature and output rule traceability statement;S6, intercept early warning before multi-modal data and traceability statement are packaged and generated rule traceability evidence package.The application realizes dynamic scene self-adapting compliance determination and explainable attribution evidence, improves early warning accuracy and supervision closed loop efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent security and police compliance technology, and in particular to a method for early warning of abnormal behavior of police inspectors based on behavioral feature recognition. Background Technology

[0002] Technologies based on multimodal fusion and explainable artificial intelligence, possessing multidimensional perception and logical reasoning capabilities in complex scenarios, have been widely applied in recent years in fields such as intelligent security, police supervision, and compliance review, becoming an important development direction for intelligent compliance supervision. However, in practical applications, police compliance early warning scenarios face numerous challenges, including variable behavioral postures, frequent occlusion, and large differences in scene tolerance. The deployment effectiveness of multimodal behavior recognition and compliance judgment is still constrained by many factors.

[0003] Currently, most police early warning methods rely on feature input from a single visual modality, making it difficult to fully utilize multi-source heterogeneous information such as body camera status, access control trajectories, and interrogation audio. This results in a lack of contextual constraints and comprehensiveness in modeling intent to violate regulations. Some systems only use fixed static thresholds to assess behavioral risk, ignoring the differentiated requirements for compliance boundary tolerance in different scenarios such as routine interrogations, emergency detention, mandatory restraints, and night duty, thus limiting the adaptive and dynamic adjustment capabilities of early warning strategies. Furthermore, the early warning logic lacks a transparent explanatory path; traditional attribution methods rely solely on classification probabilities for reverse decoding, making it difficult to accurately pinpoint the key physical features driving behavior across compliance boundaries. This fails to provide inspectors with clear decision-making basis, impacting the credibility and usability of early warning results.

[0004] Furthermore, existing compliance boundary models lack an adaptive drift mechanism when facing dynamic scenarios, failing to dynamically scale the boundary radius based on context labels, resulting in frequent false alarms in highly sensitive scenarios and false alarms in low-sensitivity scenarios. Moreover, the evidence extraction after anomaly detection is mostly a simple stacking of time period data, lacking logical binding between violation causal statements and multimodal evidence, making it difficult to form an exclusive violation causal evidence package, which seriously affects the practical value and credibility of the intelligent supervision system in real law enforcement scenarios.

[0005] Therefore, how to provide a method for early warning of abnormal behavior of police inspectors based on behavioral feature recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for early warning of abnormal behavior in police supervision based on behavioral feature recognition. This invention fully integrates key steps such as multimodal stream synchronous acquisition, compliance knowledge graph vectorization, Markov chain intent evolution prediction, boundary adaptive drift, and reverse decoding attribution. It constructs an intelligent supervision process with dynamic context state machine driving, adaptive adjustment of compliance boundaries, and precise tracing of violation features, achieving comprehensive perception, early warning, and dynamic verification of abnormal behavior in complex law enforcement scenarios. This invention introduces an improved Grad-CAM for reverse decoding attribution, constructing a gradient backpropagation target scalar with the distance from the current behavioral feature point to the compliance boundary hyperplane as the gradient, replacing the traditional classification probability target. This accurately filters out the positive contribution regions driving behavior across compliance boundaries, extracts Top-K key feature dimensions, and combines multimodal evidence to form a logically exclusive violation tracing evidence package. This invention has the advantages of adaptive scene boundary drift, strong geometric interpretation of violation attribution, and traceability of early warning results. It can significantly improve the accuracy of early warning and the credibility of violation tracing in dynamic scenarios, thereby effectively solving the problems of rigid compliance boundaries, one-sided early warning evaluation and opaque decision-making process in existing methods.

[0007] A method for early warning of abnormal behavior of police inspectors based on behavioral feature recognition according to an embodiment of the present invention includes the following steps: S1. Simultaneously collect multimodal stream data in police supervision scenarios, and perform anti-occlusion skeletal topology reconstruction on the visual stream of surveillance video, outputting behavioral feature topology sequences and multimodal stream data; S2. The preset police compliance regulations are processed into a graph to construct a compliance knowledge graph, and the compliance knowledge graph is vectorized and mapped to a compliance boundary hyperplane. S3. Based on the behavioral feature topological sequence, and combined with the state flow, trajectory flow and voiceprint flow in the multimodal flow data as temporal context constraint features, calculate the transition probability Markov chain of the current behavioral sequence to various illegal intentions, and output the early warning signal; S4. Extract state flow and trajectory flow from multimodal flow data, parse the context label of the current scene, divide the current scene state, obtain the state of normal interrogation, emergency entry, mandatory constraint or night duty, construct dynamic context state machine, drive compliance boundary hyperplane to adaptively drift, and output dynamically adjusted judgment threshold. S5. Verify the early warning signal based on the dynamically adjusted judgment threshold, introduce the improved Grad-CAM for reverse decoding and attribution, construct the gradient backpropagation target scalar with the distance from the current behavior feature point to the compliance boundary hyperplane as the gradient, calculate the gradient and weight coefficients relative to the forward propagation feature map, screen out the positive contribution area, extract the Top-K key feature dimensions that lead to the abnormal judgment, and output the violation tracing statement. S6. At the moment an abnormal warning is triggered, based on the statement of cause of violation, automatically capture the key frames of the monitoring, the footage from the law enforcement recorder and the access control records from N seconds before the warning, and package them together with the statement of cause of violation into a violation source tracing evidence package, and output it to the police supervision platform.

[0008] Optionally, S1 specifically includes: S11. Visual streams of surveillance video are collected by high-definition network cameras deployed above the case-handling area; status streams of equipment power-on / off and recording status are collected by law enforcement recorders worn by police officers; trajectory streams of personnel passage records are collected by turnstile card readers at the entrance and exit of the case-handling area; voiceprint streams of audio from the interrogation room are collected by microphones installed in the interrogation room; and synchronization trigger signals are sent to each device using a unified clock source to align the timestamps of the collected data. S12. Convert each video frame of the visual stream into an RGB color space image, extract the battery level, recording and connection status codes of the law enforcement recorder from the status stream, extract the time node and personnel identification of the access control card swipe from the trajectory stream, and perform voice endpoint detection on the voiceprint stream to extract valid audio segments and voiceprint feature vectors. S13. Use the human pose estimation model to extract the coordinates of the two-dimensional human skeleton key points from the RGB image of the visual flow, and draw the connection between adjacent key points according to the connection relationship of the human physical skeleton to construct the initial skeleton topology map. S14. For the occluded key points in the initial skeleton topology map, a spatiotemporal graph convolutional network is used to fuse the visible key point position information of adjacent frames before and after, calculate the predicted coordinates of the occluded key points and fill in the missing skeleton connections, perform anti-occlusion skeleton topology reconstruction, and output the behavior feature topology sequence. S15. Format and encapsulate the state stream, trajectory stream, and voiceprint stream data that have completed timestamp alignment, and output multimodal stream data.

[0009] Optionally, S2 specifically includes: S21. Read the preset police compliance regulations text, use the named entity recognition model to scan the text, record the identified police officer and suspect words as personnel nodes, record the interrogation room and waiting room words as area nodes, record the police equipment and restraint belt words as item nodes, and record the beating and insulting words as action nodes. S22. Calculate the distance in meters between personnel nodes and area nodes in physical space and record it as the attribute value of the distance edge. Calculate the maximum duration in minutes for an action node as specified in the compliance regulations and record it as the attribute value of the duration edge. Draw a directed line segment from the action node that occurs first to the action node that occurs later, according to the order described in the compliance regulations text, and record the directed line segment as the sequence edge. S23. Use all personnel nodes, area nodes, item nodes and action nodes as graph vertices, and use distance edges, duration edges and sequence edges as graph edges to connect the corresponding graph vertices to construct a compliance knowledge graph. S24. For each node and each edge in the compliance knowledge graph, randomly generate an initial vector. For each target node, extract the initial vectors of all neighboring nodes directly connected to the target node and the initial vectors of all connected edges. Concatenate all the extracted initial vectors into a long vector in dimensional order. Input the long vector into a fully connected layer network. Add the vector output by the fully connected layer network to the initial vector of the target node in the corresponding dimension to obtain the updated node feature vector. S25. In the compliance knowledge graph, select the node paths that contain only compliant behaviors as positive sample paths. Calculate the arithmetic mean of the updated node feature vectors of all nodes in the positive sample paths according to their dimensions and form a compliance center vector. Calculate the Euclidean distance between the updated node feature vector of the violation behavior node and the compliance center vector. Extract the maximum value among all the Euclidean distances calculated for all violation behavior nodes and record it as the boundary radius. S26. Using the coordinates of the compliance center vector in high-dimensional space as the center of the sphere and the boundary radius as the radius of the sphere, draw all points in high-dimensional space that are at a distance from the center of the sphere equal to the boundary radius. Connect all the points to form a closed multidimensional geometric surface, which serves as the compliance boundary hyperplane.

[0010] Optionally, S3 specifically includes: S31. The behavioral feature topology sequence is split into continuous single-frame skeleton topology features according to time steps. The state flow, trajectory flow and voiceprint flow in the multimodal stream data are aligned according to the same time step. The state code, personnel identification and voiceprint feature vector of each time step are extracted. The state code, personnel identification and voiceprint feature vector are concatenated with the single-frame skeleton topology features to obtain the temporal context constraint features of each time step. S32. Extract action nodes from the compliance knowledge graph, set the nodes that belong to the actions stipulated in the compliance regulations as normal behavior state nodes, set the nodes that belong to the violations as violation intent state nodes, and combine the normal behavior state nodes and the violation intent state nodes into a set of state nodes of the Markov chain. S33. Input the temporal context constraint features of the current time step into the fully connected layer and then use the softmax activation function to calculate the probability value of the current temporal context constraint features to each state node in the state node set. Arrange all the calculated probability values ​​to generate a transition probability matrix. S34. Based on the transition probability matrix, extract the probability value that the current time step is in the normal behavior state node and the next time step is to transition to the illegal intention state node. Record the corresponding probability value as the one-step transition probability. Multiply the one-step transition probabilities of multiple consecutive time steps in the order of time steps to obtain the transition probability Markov chain of the current behavior sequence evolving into various illegal intentions. S35. Compare the cumulative probability value of the transition probability Markov chain to any type of illegal intention state node with the preset intention warning threshold. When the cumulative probability value is greater than the preset intention warning threshold, generate a pre-warning signal and output it.

[0011] Optionally, S4 specifically includes: S41. Extract state stream and trajectory stream from multimodal stream data, count the access control card swipe time nodes extracted from the trajectory stream to obtain the number of people entering, extract the recording status code of the law enforcement recorder from the state stream to parse the corresponding task nature, and combine the number of people entering and the task nature into the context label of the current scene. S42. When the number of people entering and the nature of the task in the context label meet the preset first scenario condition, the current scenario state is classified as normal interrogation. When the number of people entering and the nature of the task meet the preset second scenario condition, the current scenario state is classified as emergency entry. When the number of people entering and the nature of the task meet the preset third scenario condition, the current scenario state is classified as mandatory constraint. When the time node in the context label meets the preset fourth scenario condition, the current scenario state is classified as night duty. S43. Set routine interrogation, emergency detention, compulsory restraint and night duty as four discrete state nodes, draw directed lines from one state node to another according to the scenario switching order stipulated in the police compliance regulations, and construct a dynamic context state machine. S44. Assign a preset first radius scaling factor, second radius scaling factor, third radius scaling factor and fourth radius scaling factor to routine interrogation, emergency entry, mandatory restraint and night duty respectively. Multiply the radius scaling factor corresponding to the current scene state by the boundary radius to obtain the dynamic radius of the current scene. S45. Using the coordinates of the compliance center vector in the high-dimensional space as the center of the sphere and the dynamic radius of the current scene as the radius of the sphere, redraw all points in the high-dimensional space whose distance to the center of the sphere is equal to the dynamic radius of the current scene, connect all the points to form a closed surface, and perform adaptive drift of the compliance boundary hyperplane. S46. Calculate the vertical distance between the spatial coordinate point corresponding to the compliance center vector and the closed surface after adaptive drift, and output the value of the vertical distance as the dynamically adjusted judgment threshold.

[0012] Optionally, S5 specifically includes: S51. After receiving the early warning signal, extract the temporal context constraint features of the current time step as the current behavior feature point, calculate the vertical distance between the current behavior feature point and the closed surface after adaptive drift, and compare it with the dynamically adjusted judgment threshold. When the value of the vertical distance is less than the dynamically adjusted judgment threshold, it is confirmed that the current behavior is outside the closed surface and is judged as an anomaly. S52. When an anomaly is confirmed, an improved Grad-CAM is introduced to calculate the vertical distance from the coordinate point of the current behavior feature point in the high-dimensional space to the compliance boundary hyperplane, and use it as the target scalar for gradient backpropagation of the improved Grad-CAM. S53. Obtain the forward propagation network used to generate the forward warning signal, and extract the forward propagation feature map from the last convolutional layer of the forward propagation network; S54. Starting with the gradient backpropagation target scalar, perform backpropagation, calculate the partial derivative of the gradient backpropagation target scalar with respect to each channel in the forward propagation feature map at each spatial location, and record the value of the partial derivative as the target gradient. S55. For each channel in the forward propagation feature map, extract the target gradient of the channel at all spatial locations, perform global average pooling operation on the target gradient of the channel at all spatial locations to calculate the arithmetic mean, and record the arithmetic mean as the weight coefficient of the channel. S56. Multiply the feature vector at each spatial location in the forward propagation feature map with the weight coefficient of its corresponding channel. Sum the results of multiplying all channels by the weight coefficient in the channel dimension and filter out negative values ​​through the ReLU activation function. Select the positive contribution regions that have a positive driving effect on reducing the distance from the current behavior feature point to the compliance boundary hyperplane and generate a category activation map. S57. Process the category activation map to obtain abnormal status, trajectory, voiceprint and action of law enforcement recorder. Combine the mapped abnormality types with the Top-K key feature dimensions to form text and output the statement of cause of violation.

[0013] Optionally, S57 specifically includes: S571. Sort the values ​​of all spatial locations in the category activation map from largest to smallest, and iterate through the sorted values ​​in turn. When the cumulative sum of the values ​​reaches the preset contribution coverage rate, stop the iteration. Extract the channel numbers in the forward propagation feature map corresponding to the positive contribution regions that were traversed before stopping. Arrange the extracted channel numbers in the traversal order to obtain the Top-K key feature dimensions that lead to the anomaly judgment. S572. Read the source of each key feature dimension in the Top-K key feature dimensions in the corresponding temporal context constraint features, map the key feature dimension whose source is state encoding to law enforcement recorder state anomaly, map the key feature dimension whose source is personnel identification to trajectory anomaly, map the key feature dimension whose source is voiceprint feature vector to voiceprint anomaly, and map the key feature dimension whose source is single-frame skeleton topology feature to action anomaly. Combine the mapped anomaly type with the Top-K key feature dimensions to form text, and output the violation cause statement.

[0014] Optionally, S6 specifically includes: S61. At the moment the abnormal warning is triggered, record the system timestamp that triggered the abnormal warning and read the mapped abnormal type from the violation cause statement; S62. Using the current system timestamp as the endpoint, backtrack for a preset extraction duration, extract video frames of the visual stream of the monitoring video within the preset extraction duration, extract the video frames at equal time intervals, and record the extracted video frames as monitoring key frames. S63. Using the current system timestamp as the endpoint, backtrack by the same interception time, extract the recording status code of the law enforcement recorder for the corresponding time period from the state stream of the multimodal stream data, and combine the extracted recording status code with the video frames of the visual stream within that time period to generate the law enforcement recorder screen. S64. Using the current system timestamp as the endpoint, backtrack to the same intercepted duration, extract the access control card swipe time node and personnel identifier for the corresponding time period from the trajectory stream of multimodal stream data, arrange the extracted time nodes and personnel identifiers in chronological order, and generate access control records. S65. Align and stitch together the monitoring keyframes, law enforcement recorder footage, access control records and violation tracing statements, use the abnormality types mapped in the violation tracing statements as index tags, and bind the monitoring keyframes, law enforcement recorder footage and access control records as supplementary evidence to the index tags to generate a violation tracing evidence package. S66. Send the evidence package of violations through the preset network port and output it to the police supervision platform for storage and display.

[0015] The beneficial effects of this invention are: This invention addresses the issue of missing behavioral features caused by misaligned timestamps and visual occlusion in police scenarios through unified clock source synchronization and anti-occlusion skeletal topology reconstruction. It employs a spatiotemporal graph convolutional network to fuse multi-frame information to complete the skeletal topology, combined with multimodal data formatting and encapsulation, outputting aligned behavioral feature topology sequences and multimodal streaming data. Furthermore, by using compliance knowledge graph vectorization and dynamic context state machine-driven approaches, it addresses the problem of fixed compliance judgment standards and large differences in scenario tolerance. It employs scene labeling and radius scaling factor calculation to perform adaptive drift of the compliance boundary hyperplane, outputting dynamically adjusted judgment thresholds. In the graph-based early warning phase, temporal contextual constraints are input into a Markov chain to calculate transition probabilities. Combined with dynamically adjusted judgment thresholds, the pre-warning signal is validated. Upon confirmation of anomalies, an improved Grad-CAM is introduced. This constructs a gradient backpropagation target scalar with the distance from the current behavioral feature point to the compliance boundary hyperplane as the gradient. Gradient weights are calculated, and positive contribution regions are selected. Top-K key feature dimensions are extracted to map anomaly types, and a violation attribution statement is output. Further, by backtracking and extracting pre-warning multimodal data and indexing the violation attribution statement, a logically exclusive violation attribution evidence package is generated and output to the supervision platform. Ultimately, this achieves a closed-loop intelligent supervision system encompassing comprehensive perception of abnormal behavior, pre-warning, and precise violation attribution in police supervision scenarios. This effectively improves the completeness of multimodal feature fusion, the scenario adaptability of compliance judgment, the interpretability of violation attribution, and the logical rigor of evidence collection. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for early warning of abnormal behavior of police inspectors based on behavioral feature recognition, as proposed in this invention. Figure 2 This is a flowchart of the compliance knowledge graph vectorization and compliance boundary hyperplane construction proposed in this invention; Figure 3 This is a flowchart of the reverse decoding attribution and violation tracing evidence package generation process proposed in this invention based on improved Grad-CAM. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A method for early warning of abnormal behavior by police inspectors based on behavioral feature recognition includes the following steps: S1. Simultaneously collect multimodal stream data in police supervision scenarios, including the status stream of law enforcement recorders, the trajectory stream of access control in the case handling area, and the voiceprint stream of audio from the interrogation room. Perform anti-occlusion skeletal topology reconstruction on the visual stream of the surveillance video and output behavioral feature topology sequences and multimodal stream data. S2. The preset police compliance regulations are processed into a graph, and a compliance knowledge graph is constructed with personnel, regions, items and actions as nodes and distance, duration and order as edges. The compliance knowledge graph is then vectorized and mapped to a compliance boundary hyperplane in a high-dimensional space. S3. Based on the behavioral feature topological sequence, and combined with the state flow, trajectory flow and voiceprint flow in the multimodal flow data as temporal context constraint features, calculate the transition probability Markov chain of the current behavioral sequence to various illegal intentions, and output the early warning signal; S4. Extract state flow and trajectory flow from multimodal flow data, parse the context label of the current scene, the context label includes the number of people entering and the nature of the task, divide the current scene state, and obtain the state of normal interrogation, emergency entry, mandatory restraint or night duty, construct a dynamic context state machine, drive the compliance boundary hyperplane to adaptively drift, and output the dynamically adjusted judgment threshold. S5. Verify the early warning signal based on the dynamically adjusted judgment threshold. When an anomaly is confirmed, introduce an improved Grad-CAM for reverse decoding and attribution. Construct a gradient backpropagation target scalar with the distance from the current behavior feature point to the compliance boundary hyperplane as the gradient. Calculate the gradient relative to the forward propagation feature map containing the behavior feature topology and multimodal flow data. Perform global average pooling along the channel dimension to obtain the weight coefficients. Filter out the positive contribution region and extract the Top-K key feature dimensions that lead to the anomaly judgment. Output a statement of violation attribution. S6. At the moment an abnormal warning is triggered, based on the violation cause statement, automatically capture the monitoring key frames, law enforcement recorder footage and access control records from N seconds before the warning, and package them with the violation cause statement into a violation source tracing evidence package with logical exclusivity, and output it to the police supervision platform.

[0019] This invention significantly improves the accuracy and scenario adaptability of police compliance early warning. By constructing a compliance knowledge graph and mapping it to a high-dimensional compliance boundary hyperplane, it achieves a precise mathematical expression of complex police compliance regulations. In dynamic law enforcement scenarios, a dynamic context state machine is introduced to drive the hyperplane's adaptive drift, effectively solving the problem of rigid tolerance of traditional fixed thresholds in scenarios such as night duty or emergency detention, fundamentally reducing false alarms and missed alarms. Simultaneously, based on multimodal flow and Markov chain-based intent evolution prediction, it achieves pre-judgment of violations. Most importantly, this invention uses an improved Grad-CAM for reverse decoding and attribution, using the distance from behavioral feature points to the hyperplane as the feedback target to accurately extract Top-K key feature dimensions and output a violation attribution statement, completely breaking the limitations of traditional "black box" early warning judgment and giving the system strong interpretability. Ultimately, the system automatically extracts multimodal data and packages it with causal statements to generate a logically exclusive evidence package for tracing violations, thus realizing a closed loop of law enforcement supervision from "discovering anomalies" to "locking in the cause" and then to "solidifying evidence," greatly enhancing the real-time intervention capability and disciplinary evidence collection efficiency of police inspectors.

[0020] In this embodiment, S1 specifically includes: S11. Visual streams of surveillance video are collected by high-definition network cameras deployed above the case-handling area; status streams of equipment power-on / off and recording status are collected by law enforcement recorders worn by police officers; trajectory streams of personnel passage records are collected by turnstile card readers at the entrance and exit of the case-handling area; voiceprint streams of audio from the interrogation room are collected by microphones installed in the interrogation room; and synchronization trigger signals are sent to each device using a unified clock source to align the timestamps of the collected data. S12. Convert each video frame of the visual stream into an RGB color space image, extract the battery level, recording and connection status codes of the law enforcement recorder from the status stream, extract the time node and personnel identification of the access control card swipe from the trajectory stream, and perform voice endpoint detection on the voiceprint stream to extract valid audio segments and voiceprint feature vectors. S13. Use the human pose estimation model to extract the coordinates of the two-dimensional human skeleton key points from the RGB image of the visual flow, and draw the connection between adjacent key points according to the connection relationship of the human physical skeleton to construct the initial skeleton topology map. S14. For the occluded key points in the initial skeleton topology map, a spatiotemporal graph convolutional network is used to fuse the visible key point position information of adjacent frames before and after, calculate the predicted coordinates of the occluded key points and fill in the missing skeleton connections, perform anti-occlusion skeleton topology reconstruction, and output the behavior feature topology sequence. S15. Format and encapsulate the state stream, trajectory stream, and voiceprint stream data that have completed timestamp alignment, and output multimodal stream data.

[0021] In this embodiment, S2 specifically includes: S21. Read the preset police compliance regulations text, use the named entity recognition model to scan the text, record the identified police officer and suspect words as personnel nodes, record the interrogation room and waiting room words as area nodes, record the police equipment and restraint belt words as item nodes, and record the beating and insulting words as action nodes. S22. Calculate the distance in meters between personnel nodes and area nodes in physical space and record it as the attribute value of the distance edge. Calculate the maximum duration in minutes for an action node as specified in the compliance regulations and record it as the attribute value of the duration edge. Draw a directed line segment from the action node that occurs first to the action node that occurs later, according to the order described in the compliance regulations text, and record the directed line segment as the sequence edge. S23. Use all personnel nodes, area nodes, item nodes and action nodes as graph vertices, and use distance edges, duration edges and sequence edges as graph edges to connect the corresponding graph vertices to construct a compliance knowledge graph. S24. For each node and each edge in the compliance knowledge graph, randomly generate an initial vector. For each target node, extract the initial vectors of all neighboring nodes directly connected to the target node and the initial vectors of all connected edges. Concatenate all the extracted initial vectors into a long vector in dimensional order. Input the long vector into a fully connected layer network. Add the vector output by the fully connected layer network to the initial vector of the target node in the corresponding dimension to obtain the updated node feature vector. S25. In the compliance knowledge graph, select the node paths that contain only compliant behaviors as positive sample paths. Calculate the arithmetic mean of the updated node feature vectors of all nodes in the positive sample paths according to their dimensions and form a compliance center vector. Calculate the Euclidean distance between the updated node feature vector of the violation behavior node and the compliance center vector. Extract the maximum value among all the Euclidean distances calculated for all violation behavior nodes and record it as the boundary radius. S26. Using the coordinates of the compliance center vector in high-dimensional space as the center of the sphere and the boundary radius as the radius of the sphere, draw all points in high-dimensional space that are at a distance from the center of the sphere equal to the boundary radius. Connect all the points to form a closed multidimensional geometric surface, which serves as the compliance boundary hyperplane.

[0022] In this embodiment, S3 specifically includes: S31. The behavioral feature topology sequence is split into continuous single-frame skeleton topology features according to time steps. The state flow, trajectory flow and voiceprint flow in the multimodal stream data are aligned according to the same time step. The state code, personnel identification and voiceprint feature vector of each time step are extracted. The state code, personnel identification and voiceprint feature vector are concatenated with the single-frame skeleton topology features to obtain the temporal context constraint features of each time step. S32. Extract action nodes from the compliance knowledge graph, set the nodes that belong to the actions stipulated in the compliance regulations as normal behavior state nodes, set the nodes that belong to the violations as violation intent state nodes, and combine the normal behavior state nodes and the violation intent state nodes into a set of state nodes of the Markov chain. S33. Input the temporal context constraint features of the current time step into the fully connected layer and then use the softmax activation function to calculate the probability value of the current temporal context constraint features to each state node in the state node set. Arrange all the calculated probability values ​​to generate a transition probability matrix. S34. Based on the transition probability matrix, extract the probability value that the current time step is in the normal behavior state node and the next time step is to transition to the illegal intention state node. Record the corresponding probability value as the one-step transition probability. Multiply the one-step transition probabilities of multiple consecutive time steps in the order of time steps to obtain the transition probability Markov chain of the current behavior sequence evolving into various illegal intentions. S35. Compare the cumulative probability value of the transition probability Markov chain to any type of illegal intention state node with a preset intention warning threshold. When the cumulative probability value is greater than the preset intention warning threshold, generate and output a pre-warning signal containing an illegal intention type label. The preset intention warning threshold is 0.6.

[0023] In this embodiment, S4 specifically includes: S41. Extract state stream and trajectory stream from multimodal stream data, count the access control card swipe time nodes extracted from the trajectory stream to obtain the number of people entering, extract the recording status code of the law enforcement recorder from the state stream to parse the corresponding task nature, and combine the number of people entering and the task nature into the context label of the current scene. S42. When the number of people entering and the nature of the task in the context label meet the preset first scenario condition, the current scenario state is classified as normal interrogation. When the number of people entering and the nature of the task meet the preset second scenario condition, the current scenario state is classified as emergency entry. When the number of people entering and the nature of the task meet the preset third scenario condition, the current scenario state is classified as mandatory constraint. When the time node in the context label meets the preset fourth scenario condition, the current scenario state is classified as night duty. The preset first scenario condition is that the number of people entering is equal to 2 and the nature of the task is routine interrogation. The preset second scenario condition is that the number of people entering is equal to 1 and the nature of the task is sudden arrest. The preset third scenario condition is that the number of people entering is equal to 2 and the nature of the task is mandatory control. The preset fourth scenario condition is that the time node is between 0:00 and 6:00. S43. Set routine interrogation, emergency detention, compulsory restraint and night duty as four discrete state nodes, draw directed lines from one state node to another according to the scenario switching order stipulated in the police compliance regulations, and construct a dynamic context state machine. S44. Assign a preset first radius scaling factor, a second radius scaling factor, a third radius scaling factor, and a fourth radius scaling factor to routine interrogation, emergency entry, mandatory restraint, and night duty respectively. Multiply the radius scaling factor corresponding to the current scene state by the boundary radius to obtain the dynamic radius of the current scene. The preset first radius scaling factor is 1.0, the preset second radius scaling factor is 1.5, the preset third radius scaling factor is 0.6, and the preset fourth radius scaling factor is 0.8. S45. Using the coordinates of the compliance center vector in the high-dimensional space as the center of the sphere and the dynamic radius of the current scene as the radius of the sphere, redraw all points in the high-dimensional space whose distance to the center of the sphere is equal to the dynamic radius of the current scene, connect all the points to form a closed surface, and perform adaptive drift of the compliance boundary hyperplane. S46. Calculate the vertical distance between the spatial coordinate point corresponding to the compliance center vector and the closed surface after adaptive drift, and output the value of the vertical distance as the dynamically adjusted judgment threshold.

[0024] In this embodiment, S5 specifically includes: S51. After receiving the early warning signal, extract the temporal context constraint features of the current time step as the current behavior feature point, calculate the vertical distance between the current behavior feature point and the closed surface after adaptive drift, and compare it with the dynamically adjusted judgment threshold. When the value of the vertical distance is less than the dynamically adjusted judgment threshold, it is confirmed that the current behavior is outside the closed surface and is judged as an anomaly. S52. When an anomaly is confirmed, an improved Grad-CAM is introduced to calculate the vertical distance from the coordinate point of the current behavior feature point in the high-dimensional space to the compliance boundary hyperplane, and use it as the gradient backpropagation target scalar of the improved Grad-CAM, replacing the operation of using classification probability as the target in the traditional Grad-CAM. S53. Obtain the forward propagation network used to generate the forward warning signal, and extract the forward propagation feature map containing behavioral feature topology and multimodal flow data from the last convolutional layer of the forward propagation network. S54. Perform backpropagation starting from the gradient backpropagation target scalar, calculate the partial derivative of the gradient backpropagation target scalar with respect to each channel in the forward propagation feature map at each spatial location, and record the value of the partial derivative as the target gradient. The target gradient represents the contribution of each pixel in the forward propagation feature map to the distance of the current behavior feature point across the compliance boundary hyperplane. S55. For each channel in the forward propagation feature map, extract the target gradient of the channel at all spatial locations, perform global average pooling operation on the target gradient of the channel at all spatial locations to calculate the arithmetic mean, and record the arithmetic mean as the weight coefficient of the channel. S56. Multiply the feature vector at each spatial location in the forward propagation feature map with the weight coefficient of its corresponding channel. Sum the results of multiplying all channels by the weight coefficient in the channel dimension and filter out negative values ​​through the ReLU activation function. Select the positive contribution regions that have a positive driving effect on reducing the distance from the current behavior feature point to the compliance boundary hyperplane and generate a category activation map. S57. Process the category activation map to obtain abnormal status, trajectory, voiceprint and action of law enforcement recorder. Combine the mapped abnormality types with the Top-K key feature dimensions to form text and output the statement of cause of violation.

[0025] In this embodiment, S57 specifically includes: S571. Sort the values ​​of all spatial locations in the category activation map from largest to smallest, and iterate through the sorted values ​​in turn. When the cumulative sum of the values ​​reaches the proportion of the total sum of the values ​​of all spatial locations in the category activation map to a preset contribution coverage rate, stop the traversal, extract the channel numbers in the forward propagation feature map corresponding to the positive contribution regions traversed before stopping, arrange the extracted channel numbers in the traversal order, and obtain the Top-K key feature dimensions that lead to the anomaly judgment. The preset contribution coverage rate is 0.8. S572. Read the source of the temporal context constraint feature corresponding to each key feature dimension in S31 in the Top-K key feature dimensions. Map the key feature dimension whose source is state encoding to law enforcement recorder state anomaly, the key feature dimension whose source is personnel identification to trajectory anomaly, the key feature dimension whose source is voiceprint feature vector to voiceprint anomaly, and the key feature dimension whose source is single-frame skeleton topology feature to action anomaly. Combine the mapped anomaly type with the Top-K key feature dimensions to form text and output the violation cause statement.

[0026] This invention introduces an improved Grad-CAM approach combined with compliance boundary hyperplane distance constraints to achieve interpretable attribution and precise source tracing of violation characteristics in police compliance early warnings. It calculates the vertical distance from the current behavioral feature point to the adaptive drift closed surface to verify anomalies, breaking through the traditional probability dependence of classification. Using the vertical distance from the feature point to the hyperplane as the gradient backpropagation target scalar, it performs backpropagation, calculating the target gradient to characterize the contribution of each pixel to crossing the boundary. Channel weights are obtained through global average pooling and filtered using ReLU to accurately select the positive contribution regions driving behavioral transgressions and generate a mapping map. The Top-K key feature dimensions are extracted by sorting and summing the mapping map values, and the channel source is inversely mapped to state, trajectory, voiceprint, or action anomalies. This invention breaks through the limitations of the black box in early warning in complex law enforcement scenarios, significantly improving the fine-grainedness and accuracy of violation attribution, and making the early warning results logically self-consistent and interpretable, achieving precise source tracing from anomaly perception to cause identification.

[0027] The improved Grad-CAM model of this invention is similar to the original Grad-CAM model in that both retain the core architecture of gradient-weighted class activation mapping. That is, both calculate the gradient of the target scalar relative to the forward propagation feature map through backpropagation, perform global average pooling on the channel gradient to obtain the weight coefficients, and then sum the feature map with the weight coefficients in a weighted manner. After filtering out negative values ​​through the ReLU activation function, the regions that have a positive contribution to the target are selected to generate class activation maps. Both rely on gradient information to quantify the feature contribution.

[0028] The difference lies in that this invention breaks the limitation of the original Grad-CAM, which uses classification probability as the gradient backpropagation target, and constructs a dynamic distance attribution system based on compliance boundaries. Building upon the original model's direct backpropagation using classification probability scores as the target, this invention innovatively calculates the perpendicular distance from the coordinates of the current behavior feature point in high-dimensional space to the compliance boundary hyperplane in step S52, replacing the classification probability as the gradient backpropagation target scalar for the improved Grad-CAM. Subsequently, in step S54, backpropagation is performed starting from this distance scalar. The calculated target gradient represents the contribution of each pixel in the forward propagation feature map to the distance of the current behavior feature point across the compliance boundary hyperplane, rather than its contribution to the classification result. Furthermore, in step S571, by sorting the mapping map values ​​and truncating the summation based on a contribution coverage rate of 0.8, the Top-K key feature dimensions are accurately extracted, and in step S572, these are inversely mapped to specific abnormal text types such as actions, trajectories, and voiceprints.

[0029] Based on the above improvements, the beneficial effects of this invention are that by replacing classification probability with the vertical distance from a point to a hyperplane, the improved Grad-CAM can directly quantify the actual degree to which each feature contributes to the behavioral boundary crossing process, completely overcoming the semantic ambiguity and gradient saturation defects of traditional probabilistic targets in compliance judgment; this design significantly enhances the physical interpretability and fine granularity of attribution, enabling the warning not only to locate violations, but also to accurately analyze which modality crossed the boundary; combined with the cumulative coverage truncation and source mapping mechanism, it achieves accurate source tracing from abstract features to concrete violation types, giving the system extremely strong logical self-consistency and disciplinary evidence collection efficiency.

[0030] In this embodiment, S6 specifically includes: S61. At the moment the abnormal warning is triggered, record the system timestamp that triggered the abnormal warning and read the mapped abnormal type from the violation cause statement; S62. Using the current system timestamp as the endpoint, backtrack for a preset extraction time, extract video frames of the visual stream of the monitoring video within the preset extraction time, extract the video frames at equal time intervals, and record the extracted video frames as monitoring key frames. The preset extraction time is 30 seconds. S63. Using the current system timestamp as the endpoint, backtrack by the same interception time, extract the recording status code of the law enforcement recorder for the corresponding time period from the state stream of the multimodal stream data, and combine the extracted recording status code with the video frames of the visual stream within that time period to generate the law enforcement recorder screen. S64. Using the current system timestamp as the endpoint, backtrack to the same intercepted duration, extract the access control card swipe time node and personnel identifier for the corresponding time period from the trajectory stream of multimodal stream data, arrange the extracted time nodes and personnel identifiers in chronological order, and generate access control records. S65. Align and stitch together the monitoring keyframes, law enforcement recorder footage, access control records and violation tracing statements, use the abnormality types mapped in the violation tracing statements as index tags, and bind the monitoring keyframes, law enforcement recorder footage and access control records as supplementary evidence to the index tags to generate a violation tracing evidence package with logical exclusivity. S66. Send the evidence package for tracing the source of violations through a preset network port and output it to the police supervision platform for storage and display. The preset network port is port 8080.

[0031] Example 1: To verify the feasibility of this invention in the field of intelligent supervision of police law enforcement compliance, it was deployed in the intelligent supervision platform of a municipal public security bureau in a large city in East my country. This bureau has 12 grassroots police stations and case-handling centers, covering 48 key monitoring points including interrogation rooms, waiting rooms, and case-handling area corridors. It receives an average of over 850 on-duty police officers and auxiliary law enforcement personnel daily, operating 24 / 7. The compliance risk events faced by grassroots police law enforcement scenarios are complex and have serious consequences. Typical events include: unsupervised individuals leaving their posts during nighttime interrogations, intentional shutdown or disconnection of body cameras, unauthorized use of police equipment in case-handling areas, and abnormal restraint postures of suspects. The platform receives over 2.4TB of behavioral data daily, including high-definition surveillance video streams, body camera status streams, case-handling area access control trajectory streams, and interrogation room microphone voiceprint streams. Traditional police supervision systems mainly rely on manual patrols and simple offline video blackout detection, resulting in severely delayed responses and failing to meet the stringent requirements of real-time early warning and closed-loop supervision of the entire law enforcement process.

[0032] In practical deployment, the method of this invention transforms multi-source heterogeneous police data into time-series structured input, constructing a police compliance knowledge graph and a dynamic context state machine. Single-frame skeletal topology nodes, voiceprint feature nodes, access control trajectory nodes, and law enforcement recorder state nodes are vectorized, and associated edges are constructed by fusing physical distance in the interrogation room, action duration, and the order of compliance regulations, forming a multimodal dynamic behavioral state graph. The temporal context constraint feature subgraph of the current law enforcement scenario is extracted and input into a Markov chain intent evolution prediction model to dynamically calculate the behavior transition probability. For different scenarios, an adaptive drift of the execution compliance boundary hyperplane is calculated using a radius scaling factor to achieve dynamic scaling of the judgment threshold. Once an anomaly is confirmed to trigger an early warning, an improved Grad-CAM is introduced for reverse decoding and attribution. A gradient backpropagation target scalar is constructed using the distance from the current behavioral feature point to the compliance boundary hyperplane as the gradient. The Top-K key feature dimensions driving the behavior out of bounds are extracted, mapped to specific violation types, and a violation tracing statement is output. Furthermore, monitoring keyframes, access control records, and device status from the 30 seconds prior to the warning are automatically extracted to generate an exclusive violation tracing evidence package. Table 1 below shows detailed comparative data between the method of this invention and existing traditional inspection methods during a three-month practical exercise and trial run. Table 1. Multi-dimensional performance comparison analysis of the present invention and traditional methods

[0033] Based on the comparative data shown in Table 1, it can be seen that the early warning method based on multimodal perception and reverse decoding attribution proposed in this invention has significant performance advantages over traditional rule-based methods in the identification of dynamic compliance in law enforcement. In particular, it has achieved comprehensive improvement in key indicators such as early warning accuracy, level matching degree, false alarm and false alarm control, and evidence chain tracing completeness.

[0034] Regarding the accuracy of early warnings, this invention maintains a high level of over 94% in all five typical police risk scenarios, far exceeding the average accuracy of traditional systems (approximately 67%). For example, in the scenario of "suspect being abnormally over-restrained," traditional systems rely on simple video motion detection and fixed posture comparison, which easily confuses the complex actions of officers in normal uniforms with those of unlawful restraints, achieving only 65.4% accuracy. In contrast, this invention, through multimodal streaming data fusion and temporal contextual constraint feature extraction, accurately distinguishes action intent, achieving 94.0% accuracy, effectively improving the ability to identify highly sensitive law enforcement behaviors.

[0035] Regarding risk level matching accuracy, this invention maintains a success rate above 89% across all scenarios, while traditional methods typically hover around 50%. Traditional methods use fixed thresholds to classify risk levels, which often leads to an imbalance in tolerance levels in dynamic scenarios such as nighttime fatigue or emergency entry. This invention, however, utilizes a dynamic context state machine to drive adaptive drift of the compliance boundary hyperplane, ensuring the accuracy and differentiation of subsequent response strategies.

[0036] The invention also demonstrates significant advantages in controlling false alarm and false negative rates. The average false alarm rate is controlled below 4%, and the false negative rate is around 3.5%. Compared to the traditional method's nearly 16% false alarm rate and approximately 11% false negative rate, this significantly reduces redundant warnings and omissions, lowers manual verification costs, and improves the stability of the monitoring system.

[0037] Its advantages are particularly prominent in terms of the completeness of the evidence chain tracing. The average completeness of the tracing is over 95%, while traditional methods are less than 30%. Traditional early warnings only output alarm signals and lack interpretive logic, while this invention introduces an improved Grad-CAM to calculate the target scalar gradient, extracts Top-K feature dimensions to accurately map anomaly types, and automatically packages multimodal data before the early warning to generate an exclusive evidence package for tracing violations, completely solving the pain points of traditional early warnings being uninterpretable and unable to serve as a basis for disciplinary action.

[0038] Overall, this invention achieves efficient, accurate, and traceable compliance risk identification and early warning in key law enforcement scenarios such as "absence of single-person supervision at night" and "unauthorized physical contact," through dynamic graph construction, adaptive boundary drift, and interpretable attribution mechanisms, thus filling the technical shortcomings of traditional methods in dynamic law enforcement risk control.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of abnormal behavior by police inspectors based on behavioral feature recognition, characterized in that, Includes the following steps: S1. Simultaneously collect multimodal stream data in police supervision scenarios, and perform anti-occlusion skeletal topology reconstruction on the visual stream of surveillance video, outputting behavioral feature topology sequences and multimodal stream data; S2. The preset police compliance regulations are processed into a graph to construct a compliance knowledge graph, and the compliance knowledge graph is vectorized and mapped to a compliance boundary hyperplane. S3. Based on the behavioral feature topological sequence, and combined with the state flow, trajectory flow and voiceprint flow in the multimodal flow data as temporal context constraint features, calculate the transition probability Markov chain of the current behavioral sequence to various illegal intentions, and output the early warning signal; S4. Extract state flow and trajectory flow from multimodal flow data, parse the context label of the current scene, divide the current scene state, obtain the state of normal interrogation, emergency entry, mandatory constraint or night duty, construct dynamic context state machine, drive compliance boundary hyperplane to adaptively drift, and output dynamically adjusted judgment threshold. S5. Verify the early warning signal based on the dynamically adjusted judgment threshold, introduce the improved Grad-CAM for reverse decoding and attribution, construct the gradient backpropagation target scalar with the distance from the current behavior feature point to the compliance boundary hyperplane as the gradient, calculate the gradient and weight coefficients relative to the forward propagation feature map, screen out the positive contribution area, extract the Top-K key feature dimensions that lead to the abnormal judgment, and output the violation tracing statement. S6. At the moment an abnormal warning is triggered, based on the statement of cause of violation, automatically capture the key frames of the monitoring, the footage from the law enforcement recorder and the access control records from N seconds before the warning, and package them together with the statement of cause of violation into a violation source tracing evidence package, and output it to the police supervision platform.

2. The method for early warning of abnormal behavior of police inspectors based on behavioral feature recognition according to claim 1, characterized in that, S1 specifically includes: S11. Visual streams of surveillance video are collected by high-definition network cameras deployed above the case-handling area; status streams of equipment power-on / off and recording status are collected by law enforcement recorders worn by police officers; trajectory streams of personnel passage records are collected by turnstile card readers at the entrance and exit of the case-handling area; voiceprint streams of audio from the interrogation room are collected by microphones installed in the interrogation room; and synchronization trigger signals are sent to each device using a unified clock source to align the timestamps of the collected data. S12. Convert each video frame of the visual stream into an RGB color space image, extract the battery level, recording and connection status codes of the law enforcement recorder from the status stream, extract the time node and personnel identification of the access control card swipe from the trajectory stream, and perform voice endpoint detection on the voiceprint stream to extract valid audio segments and voiceprint feature vectors. S13. Use the human pose estimation model to extract the coordinates of the two-dimensional human skeleton key points from the RGB image of the visual flow, and draw the connection between adjacent key points according to the connection relationship of the human physical skeleton to construct the initial skeleton topology map. S14. For the occluded key points in the initial skeleton topology map, a spatiotemporal graph convolutional network is used to fuse the visible key point position information of adjacent frames before and after, calculate the predicted coordinates of the occluded key points and fill in the missing skeleton connections, perform anti-occlusion skeleton topology reconstruction, and output the behavior feature topology sequence. S15. Format and encapsulate the state stream, trajectory stream, and voiceprint stream data that have completed timestamp alignment, and output multimodal stream data.

3. The method for early warning of abnormal behavior of police inspectors based on behavioral feature recognition according to claim 1, characterized in that, S2 specifically includes: S21. Read the preset police compliance regulations text, use the named entity recognition model to scan the text, record the identified police officer and suspect words as personnel nodes, record the interrogation room and waiting room words as area nodes, record the police equipment and restraint belt words as item nodes, and record the beating and insulting words as action nodes. S22. Calculate the distance in meters between personnel nodes and area nodes in physical space and record it as the attribute value of the distance edge. Calculate the maximum duration in minutes for an action node as specified in the compliance regulations and record it as the attribute value of the duration edge. Draw a directed line segment from the action node that occurs first to the action node that occurs later, according to the order described in the compliance regulations text, and record the directed line segment as the sequence edge. S23. Use all personnel nodes, area nodes, item nodes and action nodes as graph vertices, and use distance edges, duration edges and sequence edges as graph edges to connect the corresponding graph vertices to construct a compliance knowledge graph. S24. For each node and each edge in the compliance knowledge graph, randomly generate an initial vector. For each target node, extract the initial vectors of all neighboring nodes directly connected to the target node and the initial vectors of all connected edges. Concatenate all the extracted initial vectors into a long vector in dimensional order. Input the long vector into a fully connected layer network. Add the vector output by the fully connected layer network to the initial vector of the target node in the corresponding dimension to obtain the updated node feature vector. S25. In the compliance knowledge graph, select the node paths that contain only compliant behaviors as positive sample paths. Calculate the arithmetic mean of the updated node feature vectors of all nodes in the positive sample paths according to their dimensions and form a compliance center vector. Calculate the Euclidean distance between the updated node feature vector of the violation behavior node and the compliance center vector. Extract the maximum value among all the Euclidean distances calculated for all violation behavior nodes and record it as the boundary radius. S26. Using the coordinates of the compliance center vector in high-dimensional space as the center of the sphere and the boundary radius as the radius of the sphere, draw all points in high-dimensional space that are at a distance from the center of the sphere equal to the boundary radius. Connect all the points to form a closed multidimensional geometric surface, which serves as the compliance boundary hyperplane.

4. The method for early warning of abnormal behavior of police inspectors based on behavioral feature recognition according to claim 1, characterized in that, S3 specifically includes: S31. The behavioral feature topology sequence is split into continuous single-frame skeleton topology features according to time steps. The state flow, trajectory flow and voiceprint flow in the multimodal stream data are aligned according to the same time step. The state code, personnel identification and voiceprint feature vector of each time step are extracted. The state code, personnel identification and voiceprint feature vector are concatenated with the single-frame skeleton topology features to obtain the temporal context constraint features of each time step. S32. Extract action nodes from the compliance knowledge graph, set the nodes that belong to the actions stipulated in the compliance regulations as normal behavior state nodes, set the nodes that belong to the violations as violation intent state nodes, and combine the normal behavior state nodes and the violation intent state nodes into a set of state nodes of the Markov chain. S33. Input the temporal context constraint features of the current time step into the fully connected layer and then use the softmax activation function to calculate the probability value of the current temporal context constraint features to each state node in the state node set. Arrange all the calculated probability values ​​to generate a transition probability matrix. S34. Based on the transition probability matrix, extract the probability value that the current time step is in the normal behavior state node and the next time step is to transition to the illegal intention state node. Record the corresponding probability value as the one-step transition probability. Multiply the one-step transition probabilities of multiple consecutive time steps in the order of time steps to obtain the transition probability Markov chain of the current behavior sequence evolving into various illegal intentions. S35. Compare the cumulative probability value of the transition probability Markov chain to any type of illegal intention state node with the preset intention warning threshold. When the cumulative probability value is greater than the preset intention warning threshold, generate a pre-warning signal and output it.

5. The method for early warning of abnormal behavior of police inspectors based on behavioral feature recognition according to claim 1, characterized in that, S4 specifically includes: S41. Extract state stream and trajectory stream from multimodal stream data, count the access control card swipe time nodes extracted from the trajectory stream to obtain the number of people entering, extract the recording status code of the law enforcement recorder from the state stream to parse the corresponding task nature, and combine the number of people entering and the task nature into the context label of the current scene. S42. When the number of people entering and the nature of the task in the context label meet the preset first scenario condition, the current scenario state is classified as normal interrogation. When the number of people entering and the nature of the task meet the preset second scenario condition, the current scenario state is classified as emergency entry. When the number of people entering and the nature of the task meet the preset third scenario condition, the current scenario state is classified as mandatory constraint. When the time node in the context label meets the preset fourth scenario condition, the current scenario state is classified as night duty. S43. Set routine interrogation, emergency detention, compulsory restraint and night duty as four discrete state nodes, draw directed lines from one state node to another according to the scenario switching order stipulated in the police compliance regulations, and construct a dynamic context state machine. S44. Assign a preset first radius scaling factor, second radius scaling factor, third radius scaling factor and fourth radius scaling factor to routine interrogation, emergency entry, mandatory restraint and night duty respectively. Multiply the radius scaling factor corresponding to the current scene state by the boundary radius to obtain the dynamic radius of the current scene. S45. Using the coordinates of the compliance center vector in the high-dimensional space as the center of the sphere and the dynamic radius of the current scene as the radius of the sphere, redraw all points in the high-dimensional space whose distance to the center of the sphere is equal to the dynamic radius of the current scene, connect all the points to form a closed surface, and perform adaptive drift of the compliance boundary hyperplane. S46. Calculate the vertical distance between the spatial coordinate point corresponding to the compliance center vector and the closed surface after adaptive drift, and output the value of the vertical distance as the dynamically adjusted judgment threshold.

6. The method for early warning of abnormal behavior of police inspectors based on behavioral feature recognition according to claim 1, characterized in that, S5 specifically includes: S51. After receiving the early warning signal, extract the temporal context constraint features of the current time step as the current behavior feature point, calculate the vertical distance between the current behavior feature point and the closed surface after adaptive drift, and compare it with the dynamically adjusted judgment threshold. When the value of the vertical distance is less than the dynamically adjusted judgment threshold, it is confirmed that the current behavior is outside the closed surface and is judged as an anomaly. S52. When an anomaly is confirmed, an improved Grad-CAM is introduced to calculate the vertical distance from the coordinate point of the current behavior feature point in the high-dimensional space to the compliance boundary hyperplane, and use it as the target scalar for gradient backpropagation of the improved Grad-CAM. S53. Obtain the forward propagation network used to generate the forward warning signal, and extract the forward propagation feature map from the last convolutional layer of the forward propagation network; S54. Starting with the gradient backpropagation target scalar, perform backpropagation, calculate the partial derivative of the gradient backpropagation target scalar with respect to each channel in the forward propagation feature map at each spatial location, and record the value of the partial derivative as the target gradient. S55. For each channel in the forward propagation feature map, extract the target gradient of the channel at all spatial locations, perform global average pooling operation on the target gradient of the channel at all spatial locations to calculate the arithmetic mean, and record the arithmetic mean as the weight coefficient of the channel. S56. Multiply the feature vector at each spatial location in the forward propagation feature map with the weight coefficient of its corresponding channel. Sum the results of multiplying all channels by the weight coefficient in the channel dimension and filter out negative values ​​through the ReLU activation function. Select the positive contribution regions that have a positive driving effect on reducing the distance from the current behavior feature point to the compliance boundary hyperplane and generate a category activation map. S57. Process the category activation map to obtain abnormal status, trajectory, voiceprint and action of law enforcement recorder. Combine the mapped abnormality types with the Top-K key feature dimensions to form text and output the statement of cause of violation.

7. A method for early warning of abnormal behavior of police inspectors based on behavioral feature recognition according to claim 6, characterized in that, Specifically, S57 includes: S571. Sort the values ​​of all spatial locations in the category activation map from largest to smallest, and iterate through the sorted values ​​in turn. When the cumulative sum of the values ​​reaches the preset contribution coverage rate, stop the iteration. Extract the channel numbers in the forward propagation feature map corresponding to the positive contribution regions that were traversed before stopping. Arrange the extracted channel numbers in the traversal order to obtain the Top-K key feature dimensions that lead to the anomaly judgment. S572. Read the source of each key feature dimension in the Top-K key feature dimensions in the corresponding temporal context constraint features, map the key feature dimension whose source is state encoding to law enforcement recorder state anomaly, map the key feature dimension whose source is personnel identification to trajectory anomaly, map the key feature dimension whose source is voiceprint feature vector to voiceprint anomaly, and map the key feature dimension whose source is single-frame skeleton topology feature to action anomaly. Combine the mapped anomaly type with the Top-K key feature dimensions to form text, and output the violation cause statement.

8. The method for early warning of abnormal behavior of police inspectors based on behavioral feature recognition according to claim 1, characterized in that, S6 specifically includes: S61. At the moment the abnormal warning is triggered, record the system timestamp that triggered the abnormal warning and read the mapped abnormal type from the violation cause statement; S62. Using the current system timestamp as the endpoint, backtrack for a preset extraction duration, extract video frames of the visual stream of the monitoring video within the preset extraction duration, extract the video frames at equal time intervals, and record the extracted video frames as monitoring key frames. S63. Using the current system timestamp as the endpoint, backtrack by the same interception time, extract the recording status code of the law enforcement recorder for the corresponding time period from the state stream of the multimodal stream data, and combine the extracted recording status code with the video frames of the visual stream within that time period to generate the law enforcement recorder screen. S64. Using the current system timestamp as the endpoint, backtrack to the same intercepted duration, extract the access control card swipe time node and personnel identifier for the corresponding time period from the trajectory stream of multimodal stream data, arrange the extracted time nodes and personnel identifiers in chronological order, and generate access control records. S65. Align and stitch together the monitoring keyframes, law enforcement recorder footage, access control records and violation tracing statements, use the abnormality types mapped in the violation tracing statements as index tags, and bind the monitoring keyframes, law enforcement recorder footage and access control records as supplementary evidence to the index tags to generate a violation tracing evidence package. S66. Send the evidence package of violations through the preset network port and output it to the police supervision platform for storage and display.

Images