Security monitoring and abnormality identification method and system based on computer vision

Abstract

The application discloses a security monitoring anomaly recognition method based on computer vision, comprising the following steps: collecting continuous monitoring video, establishing visual interaction relationship, and obtaining continuous visual interaction state sequence; constructing a double-role visual relationship graph sequence; calling a GraphScope dynamic graph mining algorithm based on the double-role visual relationship graph sequence, and generating a mirror encoding state sequence; performing GraphScope state attribution encoding competition, and obtaining a graph segment interpretation state; performing local interaction structure stripping, and determining a minimum mismatch block state; reversely adjusting a role partition and a mirror residual error coding relationship, and generating an abnormal visual interaction state; and reflecting the abnormal visual interaction state to the continuous monitoring video, and outputting a security monitoring anomaly recognition result. The application improves the recognition accuracy and local anomaly positioning precision of abnormal visual interaction in a complex dynamic security scene.

Description

Technical Field

[0001] This invention relates to the field of intelligent security technology, and in particular to a method and system for anomaly identification in security monitoring based on computer vision. Background Technology

[0002] With the development of urban public safety management and video surveillance infrastructure, computer vision technology is increasingly being applied to the automatic identification of crowd gatherings, abnormal movements, target loitering, chases and conflicts, and area intrusions. Existing security monitoring anomaly identification methods typically first perform target detection and trajectory tracking on continuous monitoring video, then extract behavioral features from target position, speed, direction of movement, distance between targets, and local temporal changes, and use classification models, temporal models, or graph structure models to determine whether anomalies have occurred in the monitored scene. Some dynamic graph analysis methods also organize targets and their interactions into dynamic graphs, identifying behavioral state transitions through graph structure changes or encoding cost changes, thereby reducing the workload of continuously reviewing monitoring videos manually.

[0003] However, target interactions in security monitoring scenarios exhibit a clear directionality. The same monitored target may act as both the initiator and the recipient of an interaction. Simply using dynamic graph partitioning based on changes in target node connections makes it difficult to simultaneously preserve the identity associations of the same target in both interactive roles and the asymmetric changes in interactions. When changes occur in pedestrian density, movement direction, or scene activity cycles, previously observed normal behavioral states may reappear. Existing dynamic graph segmentation methods are prone to misjudging state recurrences as new abnormal states, thereby increasing false alarms caused by changes in normal scene conditions.

[0004] In addition, existing abnormal graph structure recognition usually uses changes in the structure of the entire graph or the overall abnormal score as the basis for judgment. When a small number of local abnormal interactions are embedded in a large range of normal visual relationships, it is difficult to accurately determine the smallest local structure that actually causes the state change. The abnormal part may also affect the partitioning results of the normal background relationship.

[0005] Therefore, how to provide a computer vision-based method and system for anomaly identification in security monitoring 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 computer vision-based method and system for anomaly identification in security monitoring. By introducing role mirroring relationships, GraphScope state attribution coding competition, and minimum mismatch block identification, this invention effectively distinguishes stable behaviors from abnormal interactions in dynamic monitoring scenarios, reduces misjudgments caused by changes in normal behavior, and improves the accuracy of identifying abnormal visual interactions, the precision of local anomaly localization, and the reliability of anomaly behavior determination in complex security scenarios.

[0007] The security monitoring anomaly identification method based on computer vision according to an embodiment of the present invention includes the following steps: Collect continuous monitoring video, perform visual positioning and tracking of the monitored targets, establish visual interaction relationships, and obtain a continuous visual interaction state sequence; The continuous visual interaction state sequence is converted into a visual relationship structure with a two-sided role correspondence, and a two-role visual relationship graph sequence is constructed. Based on the dual-role visual relationship graph sequence, the GraphScope dynamic graph mining algorithm is called to establish the role mirror correspondence, and the graph structure encoding change and mirror residual encoding change are jointly included in the partition search to generate the mirror encoding state sequence. For mirror-encoded state sequences, GraphScope state attribution encoding competition is performed. The behavioral state attribution is determined based on the visual interaction description cost, and the graph segment interpretation state is obtained. Based on the new behavioral state in the diagram interpretation state, perform local interaction structure stripping, and determine the minimum mismatch block state based on the condition that the recurrent encoding after stripping gains an advantage and loses its advantage after further reduction. Based on the minimum mismatch block state, restore the background visual relationship to a stable behavior state, retain the minimum mismatch block in the new behavior state, reverse the relationship between the role partition and the mirror residual encoding, and generate an abnormal visual interaction state. The abnormal visual interaction status is reflected back to the continuous monitoring video to determine the abnormal monitoring behavior and output the security monitoring anomaly identification result.

[0008] Optionally, obtaining the continuous visual interaction state sequence specifically includes: Visual positioning and continuous tracking of monitoring targets in continuous monitoring videos are performed according to the video time sequence. The spatial position and motion changes of the same monitoring target at continuous monitoring time are associated to obtain the target visual motion state sequence. Establish motion relationships between monitored targets in the target visual motion state sequence, continuously analyze the relative position changes, motion direction and relationship maintenance between targets, determine the visual interaction relationship with continuous motion relationship, and obtain the visual interaction relationship sequence; The visual interaction relationship sequence is state-sequenced according to the video time process, recording the establishment, maintenance, direction change and release of visual interaction relationships, and continuously organizing the interaction states corresponding to the same target identity to obtain a continuous visual interaction state sequence.

[0009] Optionally, obtaining the dual-role visual relationship diagram sequence specifically includes: Read the continuous visual interaction state sequence, determine the direction of action in the visual interaction relationship at each time step, establish a two-sided role mapping for the same monitoring target according to the starting position and pointing position of the visual interaction relationship, and maintain a fixed correspondence between the two-sided roles and the identity of the monitoring target to obtain the two-sided role state. For dual-role states, the visual interaction relationships in the continuous visual interaction state sequence are converted into role interaction relationships from the interaction initiator to the interaction recipient according to the direction of action. The role nodes of the interaction initiator and the interaction recipient are organized separately, and the corresponding role nodes are connected according to the role interaction relationship to construct a dual-role visual relationship diagram at each monitoring time. The dual-role visual relationship diagrams at each monitoring moment are organized continuously according to the time sequence of the continuous visual interaction state sequence. The identity indexes of the two sides corresponding to the same monitoring target are kept continuous in adjacent monitoring moments. The role interaction relationship between the two sides is updated according to the changes in the visual interaction state. The dual-role visual relationship diagrams at continuous monitoring moments are then linked sequentially to obtain the dual-role visual relationship diagram sequence.

[0010] Optionally, obtaining the mirror-encoded state sequence specifically includes: Read the visual relationship graph sequence of the two roles, establish source role set and destination role set according to the interaction initiator side and the interaction receiver side respectively, call the GraphScope dynamic graph mining algorithm to perform initial partitioning on the source role set and destination role set, establish role mirror correspondence according to the identity correspondence of the same monitoring target in the two roles, establish mirror partition correspondence according to the role overlap relationship between the source partition and the destination partition with the same monitoring target identity, and obtain the role mirror partition status; The graph structure encoding changes of the two-sided partitions are determined based on the character mirror partition state, and the difference in graph structure encoding before and after the character partition adjustment is used as the structural encoding basis for GraphScope partition search to obtain the graph structure encoding state. Based on the graph structure coding state and the role mirror partition state, the visual interaction differences between the two sides of the same monitoring target under the corresponding mirror partition are compared to determine the mirror residual coding change. The graph structure coding change and the mirror residual coding change are jointly included in the GraphScope partition evaluation to obtain the joint partition coding state. For the alternating adjustment of the dual-side role partitions in the joint partition coding state, after the role partition on one side is adjusted, the mirror residual coding evaluation of the other side is updated according to the role mirror correspondence. The dual-side role partition adjustment is constrained by the reduction of joint coding cost to obtain the mirror coding state sequence.

[0011] Optionally, obtaining the interpretation state of the figure segment specifically includes: Read the mirrored coding state sequence, compare the bilateral role partitions, role mirror correspondence, mirror partition correspondence and mirror residual coding relationship in the continuous mirrored coding state according to the monitoring time order, extract the partition structure that continuously maintains the stable mirror partition correspondence and does not have continuous role migration as the stable partition skeleton, and associate the stable partition skeleton with the corresponding behavior state to obtain the stable behavior coding state. For stable behavior encoding state and current mirror encoding state, partition inheritance is performed. Based on the role mirror correspondence, the stable partition skeleton is mapped to the current dual role visual relationship graph. For roles that maintain the same mirror partition correspondence, the partition in the stable partition skeleton is continued. For roles whose mirror partition correspondence has changed, partition inheritance is removed and they participate in GraphScope repartitioning to obtain the partition inheritance state. For partition inheritance states, GraphScope state attribution encoding competition is performed. Role partition search is constrained by stable partition skeleton, and behavior state attribution is determined based on the visual interaction description cost corresponding to different state attributions. When the visual interaction description cost corresponding to the establishment of a new state is the lowest, a new behavior state is established, and regression blocking relationships are recorded at the same time. The segment attribution of the current mirror-encoded state is determined based on the behavior state attribution, and the regression blocking relationship is retained when establishing a new behavior state to obtain the segment interpretation state.

[0012] Optionally, obtaining the minimum mismatch block state specifically includes: Read the diagram segment interpretation status, locate the roles and visual interaction connections involved in the reversion blocking relationship in the current dual-role visual relationship diagram, expand the local structure boundary along the directly connected visual interaction connections with the reversion blocking relationship as the starting position, and determine the local interaction structure stripping range. Regression-oriented stripping is performed on the local interaction structure stripping range. Starting from the visual interaction connection involved in the regression blocking relationship, the local interaction structure is gradually removed. The stopping condition is that the visual interaction description cost of the stable state regression is lower than the visual interaction description cost of maintaining the new behavior state for the first time, and a regressible stripping block is obtained. For reversible stripping blocks, reverse structural reduction is performed to gradually restore the visual interaction connection of the reversible stripping block boundary. The minimum mismatch block is determined based on the condition that the reversible encoding after stripping gains an advantage and the advantage is lost after further reduction. The minimum mismatch block is restored to its original structural position in the current dual-role visual relationship diagram, the association between the minimum mismatch block and the new behavioral state is maintained, and the structural boundary between the minimum mismatch block and the background visual relationship that can be explained by the stable behavioral state is recorded to obtain the minimum mismatch block state.

[0013] Optionally, obtaining the abnormal visual interaction state specifically includes: Read the state of the minimum mismatch block, maintain the structural position of the minimum mismatch block in the current dual-role visual relationship graph, and restore the visual interaction relationships outside the minimum mismatch block to the stable behavior state according to the stable partition skeleton; The minimum mismatch block is kept in the new behavioral state, and the original visual interaction connection between the minimum mismatch block and the background visual relationship is maintained. For the minimum mismatch block that remains in the new behavioral state, the adjustment range is limited by the roles involved in the minimum mismatch block and the visual interaction relationships directly connected with the background visual relationship. The graph structure encoding changes and mirror residual encoding changes are re-incorporated into the GraphScope partition evaluation, and the bilateral role partition and mirror residual encoding relationship are adjusted in reverse. Based on the adjusted bilateral role partitioning and mirror residual coding relationship, the minimum mismatch block and behavior state affiliation are re-determined, and the reverse adjustment is iteratively performed until the background visual relationship maintains a stable behavior state and the minimum mismatch block maintains the new behavior state, generating an abnormal visual interaction state.

[0014] Optionally, obtaining the anomaly identification result of the security monitoring specifically includes: Read the abnormal visual interaction status, and based on the identity and time location of the monitoring target corresponding to the smallest mismatch block, replay the continuous monitoring video to determine the range of abnormal visual interaction video; To recover the monitoring target and visual interaction relationship within the range of abnormal visual interaction videos, the continuous movement process of the monitoring target and the direction of visual interaction are correlated to determine abnormal monitoring behavior; Based on the time period of abnormal monitoring behavior, the monitoring target, and the visual interaction range, the identification information is organized and the security monitoring anomaly identification results are output.

[0015] A computer vision-based security monitoring anomaly identification system according to an embodiment of the present invention includes: The visual interaction construction module is used to acquire continuous monitoring video, perform visual positioning and tracking of monitoring targets, establish visual interaction relationships, and obtain a continuous visual interaction state sequence. The dual-role composition module is used to convert a continuous sequence of visual interaction states into a visual relationship structure with corresponding roles on both sides, and to construct a sequence of dual-role visual relationship graphs. The mirror coding partitioning module is used to call the GraphScope dynamic graph mining algorithm based on the dual-role visual relationship graph sequence to establish a role mirror correspondence, and incorporate graph structure coding changes and mirror residual coding changes into the partition search to generate a mirror coding state sequence. The state attribution competition module is used to perform GraphScope state attribution encoding competition for mirror-encoded state sequences, determine the behavior state attribution based on the visual interaction description cost, and obtain the graph segment interpretation state. The minimum mismatch identification module is used to perform local interactive structure stripping based on the new behavioral state in the graph segment interpretation state. The minimum mismatch block state is determined based on the condition that the recurrent encoding after stripping gains an advantage and loses an advantage after further reduction. The abnormal interaction generation module is used to restore the background visual relationship to a stable behavior state based on the minimum mismatch block state, retain the minimum mismatch block in the new behavior state, reverse the adjustment of the role partition and mirror residual coding relationship, and generate abnormal visual interaction states. The abnormal behavior output module is used to reflect abnormal visual interaction states back to continuous monitoring video, determine abnormal monitoring behavior, and output the security monitoring anomaly identification results.

[0016] The beneficial effects of this invention are: This invention utilizes the GraphScope dynamic graph mining algorithm to perform dynamic graph structure analysis on visual interaction relationships in continuous surveillance videos. This is because abnormal behaviors in security monitoring scenarios are not merely manifested as changes in the position of a single target, but are often accompanied by the continuous establishment, maintenance, shifting, and dissolution of interaction relationships between targets over time. GraphScope can organize continuously changing visual interaction relationships into a dynamic graph and characterize the evolution of behavioral states through partitioning structure and changes in encoding cost. Compared to anomaly judgment based solely on single-frame target features or fixed time windows, it is more suitable for handling continuously changing target interaction processes in monitoring scenarios.

[0017] This invention addresses the directional nature of visual interactions in security systems by categorizing the same monitored target into interaction initiator roles and interaction receiver roles, and establishing a role mirror correspondence. This allows the original source and destination partitions of GraphScope to connect with the two behavioral roles of the monitored target. Furthermore, it incorporates graph structure encoding changes and mirror residual encoding changes into the partition evaluation, enabling role partition adjustments to not only reflect the degree of compression of the visual relationship structure but also the asymmetric changes between the two interactions of the same monitored target, thereby improving the accuracy of behavioral structure representation in complex interaction scenarios.

[0018] This invention addresses the problem that GraphScope's original dynamic graph segmentation mainly relies on changes in the current graph structure to determine state transitions, making it difficult to identify the recurrence of normal behavioral states. It introduces a stable partition skeleton and GraphScope state attribution coding competition. By competing the coding between the continuation of the current state, the return to a stable state, and the establishment of a new state, the attribution of behavioral states is determined. This allows normal visual interaction states that have formed stable interpretability in the past to be re-identified when they reappear, rather than being directly treated as new abnormal states. This reduces misjudgments caused by changes in pedestrian traffic, periodic changes in traffic patterns, and the restoration of normal interaction structures.

[0019] This invention addresses the problem of accurately locating the source of local anomalies in the overall graph structure anomaly judgment. It introduces the stripping of local interaction structures in the new behavioral state and uses the condition of gaining coding advantage after the stable state is restored after stripping and losing coding advantage after further reduction as the determination condition of the minimum mismatch block. This enables anomaly recognition to no longer be limited to the entire graph structure or the overall anomaly score, but to locate the smallest local visual interaction structure that truly causes the stable behavioral state to fail.

[0020] This invention restores the background visual relationship to a stable behavioral state, retains the minimum mismatch block in the new behavioral state, and uses the minimum mismatch block to reverse adjust the relationship between role partitioning and mirror residual encoding. This allows the determination results of abnormal local structures to continue participating in GraphScope partitioning search, forming a closed-loop correction process between anomaly localization and graph structure reconstruction. This reduces the interference of local anomalies on normal background relationships and improves the recognition accuracy of abnormal visual interactions, the localization accuracy of abnormal local structures, and the stability of abnormal monitoring behavior determination in complex dynamic security scenarios. Attached Figure Description

[0021] 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 the computer vision-based security monitoring anomaly identification method proposed in this invention; Figure 2 This diagram illustrates the abnormal interaction recognition process of the computer vision-based security monitoring anomaly recognition method proposed in this invention, which is based on GraphScope mirror encoding, state attribution competition, and minimum mismatch block feedback correction. Detailed Implementation

[0022] 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.

[0023] refer to Figures 1-2 A computer vision-based method for anomaly identification in security monitoring includes the following steps: Collect continuous monitoring video, perform visual positioning and tracking of the monitored targets, establish visual interaction relationships, and obtain a continuous visual interaction state sequence; The continuous visual interaction state sequence is converted into a visual relationship structure with a two-sided role correspondence, and a two-role visual relationship graph sequence is constructed. Based on the dual-role visual relationship graph sequence, the GraphScope dynamic graph mining algorithm is called to establish the role mirror correspondence, and the graph structure encoding change and mirror residual encoding change are jointly included in the partition search to generate the mirror encoding state sequence. For mirror-encoded state sequences, GraphScope state attribution encoding competition is performed. The behavioral state attribution is determined based on the visual interaction description cost, and the graph segment interpretation state is obtained. Based on the new behavioral state in the diagram interpretation state, perform local interaction structure stripping, and determine the minimum mismatch block state based on the condition that the recurrent encoding after stripping gains an advantage and loses its advantage after further reduction. Based on the minimum mismatch block state, restore the background visual relationship to a stable behavior state, retain the minimum mismatch block in the new behavior state, reverse the relationship between the role partition and the mirror residual encoding, and generate an abnormal visual interaction state. The abnormal visual interaction status is reflected back to the continuous monitoring video to determine the abnormal monitoring behavior and output the security monitoring anomaly identification result.

[0024] In this embodiment, obtaining the continuous visual interaction state sequence specifically includes: Visual positioning and continuous tracking of monitoring targets in continuous monitoring videos are performed according to the video time sequence. The spatial position and motion changes of the same monitoring target at continuous monitoring time are associated to obtain the target visual motion state sequence. The visual positioning refers to determining the spatial position of the monitoring target in the continuous monitoring image. The continuous tracking refers to maintaining the identity correspondence of the same monitoring target at continuous monitoring time and associating its position changes. The target visual motion state sequence is used to maintain the correspondence between the identity of the monitoring target and the continuous motion process. The identity recovery of the target after short-term occlusion is associated according to the spatial motion continuity before and after occlusion. For monitoring targets in the target visual motion state sequence, establish motion association between targets, continuously analyze the relative position changes, motion direction and relationship maintenance between targets, determine the visual interaction relationship with continuous motion association, and obtain the visual interaction relationship sequence. The visual interaction relationship refers to the directional association relationship between two monitoring targets determined by the continuous motion process, which is different from the spatial proximity relationship established only based on the target distance in a single frame. The visual interaction relationship sequence is state-sequenced according to the video time process, and the establishment, maintenance, direction change and release of the visual interaction relationship are recorded. The interaction states corresponding to the same target identity are continuously organized to obtain a continuous visual interaction state sequence. The continuous visual interaction state sequence is used to characterize the continuous state of the visual interaction relationship between the monitored targets as the monitoring time changes, and serves as the basis for constructing the dual-role visual relationship diagram sequence.

[0025] The continuous monitoring video acquisition frame rate is set to 25 frames / second, and the video resolution is set to 1920×1080 pixels. The location and identity information of the monitored target are updated every frame. During continuous tracking, the same monitored target is allowed to be temporarily occluded for a maximum of 12 frames. When the target reappears within 12 frames and the deviation of the movement direction is controlled within 15°, the original monitored target identity is restored. The visual interaction relationship is written into the visual interaction relationship sequence when it is continuously maintained for 8 frames.

[0026] In this embodiment, obtaining the dual-role visual relationship diagram sequence specifically includes: Read the continuous visual interaction state sequence, determine the direction of action in the visual interaction relationship at each time step, establish a two-sided role mapping for the same monitoring target according to the starting position and pointing position of the visual interaction relationship, and maintain a fixed correspondence between the two-sided roles and the identity of the monitoring target to obtain the two-sided role state. The two-sided role mapping is to map the monitoring target as the interaction initiator when it is at the starting position of the visual interaction, and as the interaction recipient when it is at the pointing position of the visual interaction, without changing the identity of the monitoring target. The same monitoring target can occupy two role positions as the visual interaction direction changes. The two role positions jointly retain the monitoring target identity index. The two-sided role state is used to record the interaction direction position of the monitoring target at the current monitoring time and the correspondence of the same identity, so that the identity of the monitoring target itself is separated from the role position of the monitoring target in the visual interaction. For dual-role states, the visual interaction relationships in the continuous visual interaction state sequence are converted into role interaction relationships from the interaction initiator to the interaction recipient according to the direction of action. Role nodes on the interaction initiator side and the interaction recipient side are organized separately, and corresponding role nodes are connected according to the role interaction relationship to construct a dual-role visual relationship graph for each monitoring moment. The role interaction relationship inherits the monitoring target identity, direction of action, and time position of the visual interaction relationship in the continuous visual interaction state sequence. When the same monitoring target participates in different visual interaction relationships, the corresponding role position is retained respectively, so that the same monitoring target can maintain identity association on both sides of the dual-role visual relationship graph. The dual-role visual relationship graph is a directed visual relationship structure formed by using the dual-role states as the node organization basis and the role interaction relationship as the dual-side connection basis. The interaction initiator side is the action start side determined according to the target movement direction and relative movement change in the visual interaction relationship, and the interaction recipient side is the action direction side in the visual interaction relationship, and the interaction recipient side is the corresponding action direction side, and the interaction recipient side is the corresponding action recipient side; The dual-role visual relationship graphs at each monitoring moment are organized continuously according to the time sequence of continuous visual interaction states. The identity indices of the corresponding dual-role roles for the same monitoring target are kept continuous in adjacent monitoring moments. The role interaction relationship between the dual-role roles is updated according to the changes in visual interaction state. The dual-role visual relationship graphs at consecutive monitoring moments are sequentially associated to obtain a dual-role visual relationship graph sequence. The continuous organization of identity uses the identity index of the monitoring target as the basis for the correspondence between adjacent dual-role visual relationship graphs. When the visual interaction relationship is maintained, the corresponding role nodes and role interaction relationship are continued. When the visual interaction direction changes, the position of the dual-role roles and role interaction relationship are updated. When the visual interaction relationship is terminated, the corresponding role interaction relationship is terminated. At the same time, the identity correspondence of the monitoring target in adjacent monitoring moments is retained. The dual-role visual relationship graph sequence is a set of visual relationship graphs that are continuously arranged according to the video time process and maintain the identity association between the dual-role roles of the monitoring target.

[0027] The dual-role visual relationship graph is updated once every 5 frames. Each monitored target uses a unique digital identity index. When the same monitored target is used as both the interaction initiator and the interaction recipient, the same identity index is used. When the visual interaction direction changes continuously for more than 3 frames, the positions of both sides are updated. When the visual interaction relationship is continuously terminated for more than 5 frames, the corresponding role interaction connection is deleted. This ensures that the dual-role visual relationship graph sequence maintains the continuity of the monitored target identity and the temporal consistency of the interaction direction changes.

[0028] In this embodiment, obtaining the mirror-encoded state sequence specifically includes: Read the dual-role visual relationship graph sequence, establish source-end role sets and destination-end role sets according to the interaction initiator and the interaction receiver, respectively, call the GraphScope dynamic graph mining algorithm to perform initial partitioning on the source-end role sets and destination-end role sets, establish role mirror correspondence based on the identity correspondence of the same monitoring target in the dual-side roles, establish mirror partition correspondence based on the overlap relationship of roles with the same monitoring target identity between the source-end partition and the destination partition, and obtain the role mirror partition state. The initial partitioning is to organize the source-end roles and destination-end roles according to the distribution relationship of role interaction connections in each dual-role visual relationship graph. The role mirror correspondence is to establish a dual-side identity association between the interaction initiator and the interaction receiver based on the identity identifier of the same monitoring target, which is used to maintain the identity connection of the same monitoring target in the source-end role set and the destination-end role set. The mirror partition correspondence establishes a partition association according to the monitoring target identity commonly included in the source-end partition and the destination partition, and updates it as the role partition is adjusted. The source-end partition is used to organize source-end roles with similar visual interaction connection characteristics in the interaction initiator side, and the destination-end partition is used to organize destination-end roles with similar visual interaction connection characteristics in the interaction receiver side. The graph structure encoding changes of the two-sided partitions are determined based on the role mirror partition state, and the difference in graph structure encoding before and after the role partition adjustment is used as the structural encoding basis for GraphScope partition search to obtain the graph structure encoding state. The graph structure encoding change refers to the change in the description cost of the visual relationship structure corresponding to the two-sided partitions after the role to be adjusted is migrated from the current partition to the candidate partition, relative to the state before migration. The GraphScope dynamic graph mining algorithm encodes the role partition itself and the visual interaction connection between the two-sided partitions respectively. The encoding of the role partition itself is used to describe the role's affiliation structure in the two-sided partitions, and the encoding of the visual interaction connection between the two-sided partitions is used to describe the distribution structure of the visual interaction relationship between different partitions. The reduction in graph structure encoding change indicates that the current role migration can enable the visual relationship graph of the two roles to obtain a more compact structural expression under the existing GraphScope partition structure. The graph structure encoding state is the partition search state formed after establishing the association between the role mirror partition state and the corresponding graph structure encoding change. Based on the graph structure encoding state and the role mirror partition state, the visual interaction differences between the two sides of the same monitored target under the corresponding mirror partition are compared to determine the mirror residual encoding change. The graph structure encoding change and the mirror residual encoding change are jointly incorporated into the GraphScope partition evaluation to obtain the joint partition encoding state. The visual interaction difference is determined by comparing the visual interaction relationship formed by the role on one side of the same monitored target with the visual interaction relationship formed by the other side of the role in the corresponding mirror partition of the current partition, according to the role mirror correspondence. Directional differences that cannot be mutually canceled out by role mirror correspondence and mirror partition correspondence are extracted. The mirror residual is the encoded representation of the directional difference in the current role partition structure. The role to be adjusted is... After the current partition is migrated into the candidate partition, the directional differences in the associated mirror partition are re-determined, and the mirror residual description cost before and after the role migration is compared to obtain the mirror residual coding change. The GraphScope partition evaluation refers to comparing the comprehensive changes in the graph structure description cost and the mirror residual description cost before and after the role migration for the coding results after the role is migrated into the candidate partition, and judging whether the corresponding candidate partition should be retained based on whether the comprehensive description cost is reduced. This ensures that the candidate role migration is constrained by both the degree of compression of the visual relationship graph structure of the two roles and the degree of asymmetry of the interaction between the two sides of the same monitoring target. The joint partition coding state is the partition search state formed after the graph structure coding state and the mirror residual coding change are associated. For the alternating adjustment of the joint partition coding status of the two-sided role partitions, after the role partition on one side is adjusted, the mirror residual coding evaluation of the other side role is updated according to the role mirror correspondence. The adjustment of the two-sided role partitions is constrained by the reduction of joint coding cost to obtain the mirror coding status sequence. The adjustment of the two-sided role partitions includes: sequentially testing the partition migration relationship between the source role and the source candidate partition, selecting the source candidate partition that can reduce the joint coding cost to update the source role affiliation, then redetermining the mirror partition correspondence according to the updated source role affiliation, and adjusting the destination role affiliation according to the updated mirror residual coding evaluation. After the destination role is adjusted, the mirror residual coding evaluation of the corresponding source role is updated in reverse, so that the source role partition and the destination role partition are alternately updated around the same joint coding relationship. When continuing to perform role migration cannot reduce the joint coding cost, the current two-sided role partitions are maintained. The mirror residual coding relationship is used to record the visual interaction differences between the two-sided roles of the same monitoring target that cannot be explained to each other and the corresponding partition coding situation. The GraphScope dynamic graph mining algorithm uses the minimum description length principle to encode and evaluate the node partitioning structure and inter-partition connection relationships in the time-varying graph, and adjusts the partitions where the source and destination nodes are located with the goal of reducing the cost of graph structure description. The dual-role visual relationship graph sequence is used as the time-varying graph sequence processed by the GraphScope dynamic graph mining algorithm.

[0029] This invention addresses the problem that conventional GraphScope, which relies solely on dynamic graph structure encoding changes for bilateral partitioning, fails to reflect the asymmetric features of bilateral interactions of the same target in security monitoring. It introduces role mirror correspondence and mirror residual encoding relationships, incorporating both graph structure encoding changes and mirror residual encoding changes into the partitioning evaluation. This ensures that role migration is simultaneously constrained by the degree of structural compression and the difference in interaction direction, thereby improving the accuracy of visual interaction structure representation, partitioning stability, and reliability of abnormal relationship identification in complex security scenarios.

[0030] In this embodiment, obtaining the interpretation state of the image segment specifically includes: Read the mirror-encoded state sequence, compare the bilateral role partitions, role mirror correspondence, mirror partition correspondence, and mirror residual encoding relationship in the continuous mirror-encoded state according to the monitoring time order, extract the partition structure that continuously maintains the mirror partition correspondence and does not undergo continuous role migration as a stable partition skeleton, and associate the stable partition skeleton with the corresponding behavior state to obtain the stable behavior encoding state. The stable partition skeleton is the GraphScope partition structure that maintains the stability of the source end role partition, the destination end role partition, and the mirror partition correspondence during the continuous change of the mirror-encoded state. The stable partition skeleton is not changed for role migration that only occurs in a local monitoring period and restores the original partition ownership. The stable partition skeleton is separated from the stable partition skeleton for role migration that continuously changes the mirror partition correspondence. The stable behavior encoding state is a reusable behavior state formed after establishing an association between the stable partition skeleton and the graph structure encoding state and the mirror residual encoding relationship. The behavior state refers to the visual interaction state category formed by merging the continuous mirror-encoded states interpreted by the same stable partition skeleton. For stable behavior encoding state and current mirror encoding state, partition inheritance is performed. According to the role mirror correspondence, the stable partition skeleton is mapped to the current dual-role visual relationship graph. For roles that maintain the same mirror partition correspondence, the partition in the stable partition skeleton is continued. For roles whose mirror partition correspondence has changed, the partition inheritance is removed and they participate in GraphScope repartitioning to obtain the partition inheritance state. The mapping of the stable partition skeleton to the current dual-role visual relationship graph is to establish a partition inheritance relationship between the source partition and the destination partition in the stable partition skeleton and the interaction initiator role and the interaction recipient role in the current dual-role visual relationship graph according to the role mirror correspondence, and keep the source partition corresponding to the interaction initiator side and the destination partition corresponding to the interaction recipient side. For partition inheritance states, GraphScope state attribution encoding competition is performed. Role partition search is constrained by stable partition skeletons, and the behavior state attribution is determined based on the visual interaction description cost corresponding to different state attributions. A new behavior state is established when the visual interaction description cost corresponding to the establishment of a new state is the lowest. At the same time, the regression blocking relationship is recorded. The GraphScope state attribution encoding competition includes current state continuation, stable state regression, and new state establishment. Current state continuation uses the current stable partition skeleton, stable state regression calls the historical stable partition skeleton, and new state establishment removes the inheritance restriction of the stable partition skeleton. The visual interaction description cost comprehensively represents the changes in graph structure encoding, mirror residual encoding, and additional encoding changes generated by repartitioning. The behavior state attribution is used to record the state type with the lowest visual interaction description cost. The regression blocking relationship is the local visual interaction relationship where the stable partition skeleton cannot continue to be interpreted and causes the corresponding role to remove the partition inheritance. The new behavior state is the behavior state re-established from the current mirror encoding state when the visual interaction description cost corresponding to the establishment of the new state is lower than the visual interaction description cost corresponding to the current state continuation and stable state regression in the GraphScope state attribution encoding competition. The segment affiliation of the current mirror-encoded state is determined based on the behavior state affiliation, and the regression blocking relationship is retained when establishing a new behavior state to obtain the segment interpretation state. The segment interpretation state is used to characterize the behavior state to which the current mirror-encoded state belongs and the inheritance of the stable partition skeleton, and the visual relationship range of the local interaction structure stripping is limited based on the regression blocking relationship.

[0031] This invention addresses the problem that conventional GraphScope relies solely on changes in the current graph structure to determine state transitions and struggles to identify the recurrence of stable behaviors. It introduces a stable partition skeleton and state attribution coding competition, enabling the determination of current state continuation, stable state reversion, and new state establishment under a unified visual interaction description cost. Furthermore, it preserves the reversion blocking relationship to limit the scope of abnormal structures, thereby reducing misjudgments caused by the recurrence of normal behaviors and improving the stability of behavior state determination and the reliability of anomaly identification in dynamic security scenarios.

[0032] In this embodiment, obtaining the minimum mismatch block state specifically includes: Read the diagram segment interpretation status, locate the roles and visual interaction connections involved in the regression blocking relationship in the current dual-role visual relationship diagram, extend the local structure boundary along the directly connected visual interaction connections with the regression blocking relationship as the starting position, and determine the local interaction structure stripping range. The local interaction structure stripping range is the local visual relationship area formed by continuous extension around the regression blocking relationship, which is used to limit the local interaction structure stripping to only act on the visual interaction part that prevents the stable state from reverting, and the visual interaction relationship outside the stripping range remains unchanged. Regression-oriented stripping is performed on the local interaction structure stripping range. Starting from the visual interaction connection involved in the regression blocking relationship, the local interaction structure is gradually removed. The stopping condition is that the visual interaction description cost of the stable state regression is lower than the visual interaction description cost of maintaining the new behavior state for the first time. A regressible stripping block is obtained. The regression-oriented stripping is that after each removal of a local interaction structure, GraphScope partition search is re-performed on the remaining visual interaction structure and the visual interaction description costs corresponding to the two behavior states are compared. If the visual interaction description cost of the stable state regression is still higher than the visual interaction description cost corresponding to maintaining the new behavior state, the stripping range is further expanded. Reverse structural reduction is performed on reversible stripping blocks to gradually restore the visual interaction connections at the boundaries of the reversible stripping blocks. The minimum mismatch block is determined based on the condition that the reversible encoding gains an advantage after stripping but loses its advantage after further reduction. The reverse structural reduction refers to gradually shrinking the reversible stripping blocks when the visual interaction description cost corresponding to the stable state reversal is lower than the visual interaction description cost corresponding to the new behavior state. For each restored visual interaction connection, the visual interaction description costs corresponding to the stable state reversal and the new behavior state are recompared. Reduction stops when the visual interaction description cost of the stable state reversal is no longer lower than the visual interaction description cost of the new behavior state. The minimum mismatch block is the local interaction structure that allows the stable state reversal to gain a cost advantage after stripping but loses its cost advantage after further reduction. The minimum mismatch block is restored to its original structural position in the current dual-role visual relationship graph, the association between the minimum mismatch block and the new behavioral state is maintained, and the structural boundary between the minimum mismatch block and the background visual relationship that can be explained by the stable behavioral state is recorded to obtain the minimum mismatch block state. The minimum mismatch block state is used to characterize the minimum local structure and its location that causes the current visual interaction relationship to be unable to be fully explained by the stable behavioral state, and the structural connection between the minimum mismatch block and the background visual relationship is preserved.

[0033] This invention addresses the problem that conventional anomaly graph structure analysis easily identifies overall structural changes as anomalies and struggles to pinpoint the local interactive structures that truly cause stable behavior to fail. It introduces regressive guided stripping and reverse structure reduction, and uses the occurrence of regressive coding advantages and the disappearance of advantages after further reduction to jointly limit the minimum mismatch block. This allows anomaly judgment to be shifted from the overall graph segment to the minimum causal structure, improving the accuracy of local anomaly location and reducing the interference of normal background relationships on anomaly identification results.

[0034] In this embodiment, obtaining the abnormal visual interaction state specifically includes: The minimum mismatch block state is read, and the structural position of the minimum mismatch block is maintained in the current dual-role visual relationship graph. The visual interaction relationships outside the minimum mismatch block are restored to the stable behavior state according to the stable partition skeleton. The background visual relationship is the visual interaction relationship in the current dual-role visual relationship graph that can be further explained by the stable partition skeleton after excluding the minimum mismatch block. The restoration to the stable behavior state is to redetermine the partition position of the role involved in the background visual relationship according to the stable partition skeleton, so that the background visual relationship can regain the stable state and return to the cost advantage of visual interaction description. The minimum mismatch block is kept in the new behavioral state, and the original visual interaction connection between the minimum mismatch block and the background visual relationship is maintained. This allows the background visual relationship and the minimum mismatch block to maintain a stable behavioral state and a new behavioral state, respectively. Keeping the minimum mismatch block in the new behavioral state means preserving the local interaction structure that causes the minimum mismatch block to lose the advantage of visual interaction description cost when returning to the stable state, and does not restore the minimum mismatch block to the stable behavioral state along with the background visual relationship. For the minimum mismatch block that remains in the new behavioral state, the adjustment range is limited by the roles involved in the minimum mismatch block and the visual interaction relationships directly connected to the background visual relationship. The graph structure encoding changes and mirror residual encoding changes are re-incorporated into the GraphScope partition evaluation. The bilateral role partitions and mirror residual encoding relationships are adjusted in reverse. The reverse adjustment uses the position of the minimum mismatch block to limit the roles that need to be retested for partition migration. After the role migration, the mirror partition correspondence and mirror residual encoding relationship are updated, so that the determination result of the minimum mismatch block participates in the GraphScope bilateral role partition search in reverse. The minimum mismatch block and its behavioral state affiliation are redefined based on the adjusted bilateral role partitioning and mirror residual coding relationship. The reverse adjustment is then iteratively executed until the background visual relationship maintains a stable behavioral state and the minimum mismatch block maintains a new behavioral state, generating an abnormal visual interaction state. The iterative reverse adjustment involves re-defining the role partitioning adjustment range with the updated minimum mismatch block when the redefined minimum mismatch block changes, and readjusting the bilateral role partitioning and mirror residual coding relationship again. The adjustment stops when the minimum mismatch block and its behavioral state affiliation no longer change. The abnormal visual interaction state is used to characterize the minimum mismatch block and its corresponding visual interaction structure that maintains a new behavioral state in a stable background visual relationship.

[0035] This invention addresses the problem in conventional dynamic graph anomaly recognition where local anomalies are intertwined with the stable background, and abnormal structures easily and continuously disturb the overall partitioning results. It separates and preserves the minimum mismatch block from the visual relationship with the background, and uses the minimum mismatch block to inversely constrain GraphScope role partitioning and mirror residual encoding. This allows abnormal local structures to participate in the closed-loop correction of the partitioning results, thus preserving the interpretability of stable background behavior while suppressing the spread of abnormal interactions to normal relationships. This improves the stability, boundary accuracy, and recognition reliability of abnormal visual interaction states.

[0036] In this embodiment, obtaining the anomaly identification result of the security monitoring specifically includes: Read the abnormal visual interaction status, and reflect the continuous monitoring video according to the identity and time location of the monitoring target corresponding to the smallest mismatch block to determine the range of abnormal visual interaction video. The reflection refers to locating the role in the smallest mismatch block to the corresponding monitoring target in the continuous monitoring video according to the identity of the monitoring target, and determining the corresponding continuous video time period according to the time location. To recover the monitoring target and visual interaction relationship in the abnormal visual interaction video range, the continuous motion process of the monitoring target and the visual interaction direction are associated to determine the abnormal monitoring behavior. The abnormal monitoring behavior is the actual monitoring behavior formed by the local visual interaction structure represented by the minimum mismatch block in the continuous monitoring video, and the correspondence between the abnormal visual interaction state and the original video behavior process is maintained. Based on the time period of the abnormal monitoring behavior, the monitoring target, and the visual interaction range, the identification information is organized and the security monitoring anomaly identification result is output. The security monitoring anomaly identification result is used to characterize the location, duration, monitoring target, and corresponding visual interaction range of the abnormal monitoring behavior in the continuous monitoring video.

[0037] A computer vision-based security monitoring anomaly detection system includes: The visual interaction construction module is used to acquire continuous monitoring video, perform visual positioning and tracking of monitoring targets, establish visual interaction relationships, and obtain a continuous visual interaction state sequence. The dual-role composition module is used to convert a continuous sequence of visual interaction states into a visual relationship structure with corresponding roles on both sides, and to construct a sequence of dual-role visual relationship graphs. The mirror coding partitioning module is used to call the GraphScope dynamic graph mining algorithm based on the dual-role visual relationship graph sequence to establish a role mirror correspondence, and incorporate graph structure coding changes and mirror residual coding changes into the partition search to generate a mirror coding state sequence. The state attribution competition module is used to perform GraphScope state attribution encoding competition for mirror-encoded state sequences, determine the behavior state attribution based on the visual interaction description cost, and obtain the graph segment interpretation state. The minimum mismatch identification module is used to perform local interactive structure stripping based on the new behavioral state in the graph segment interpretation state. The minimum mismatch block state is determined based on the condition that the recurrent encoding after stripping gains an advantage and loses an advantage after further reduction. The abnormal interaction generation module is used to restore the background visual relationship to a stable behavior state based on the minimum mismatch block state, retain the minimum mismatch block in the new behavior state, reverse the adjustment of the role partition and mirror residual coding relationship, and generate abnormal visual interaction states. The abnormal behavior output module is used to reflect abnormal visual interaction states back to continuous monitoring video, determine abnormal monitoring behavior, and output the security monitoring anomaly identification results.

[0038] Example 1: To verify the feasibility of the invention in practice, it was applied to a public security monitoring scenario with frequent pedestrian flow. The scenario included normal two-way traffic, periodic concentrated traffic, short stops, pedestrian convergence, reverse passage, and local crowding behavior. A small number of chasing, sudden crowding, reverse rushing into normal pedestrian flow, and local obstruction behaviors were also set to simulate the actual situation where stable background behaviors recur and local abnormal interactions are mixed with normal visual relationships. The test video resolution was 1920×1080 pixels, the acquisition frame rate was 25 frames / second, and a total of about 60 hours of continuous monitoring video and about 5.4 million frames of images were formed. After manual review, 18,430 continuous target trajectories were formed, of which 126 abnormal monitoring behaviors were marked, and the process of 92 stable behavior states reappearing was recorded.

[0039] In continuous video surveillance processing, visual positioning and continuous tracking of monitored targets are performed. The same target maintains a consistent identity across consecutive frames, allowing for short-term occlusion recovery within 12 frames. Visual interaction relationships are established when directional motion is maintained for 8 consecutive frames, and the dual-role visual relationship graph is updated at 5-frame intervals. The same monitored target retains the same identity when it is in both the interaction initiator and the interaction receiver roles. During testing, each dual-role visual relationship graph contained an average of 23.6 effective role nodes and 31.4 visual interaction connections, with a maximum of 67 effective role nodes and 103 visual interaction connections during the personnel concentration phase.

[0040] After the visual relationship graph of the two roles enters the GraphScope dynamic graph mining process, four initial partitions are set for both the source and destination role sets. Role mirroring is established based on the dual identities of the same monitoring target, and mirror partitioning is also established. Each role to be adjusted selects three candidate partitions for evaluation. Graph structure encoding changes and mirror residual encoding changes jointly participate in the GraphScope partition evaluation. The source and destination role partitions are adjusted alternately, with a maximum of 20 rounds of adjustments. The search stops when the overall descriptive cost change ratio is less than 0.5% for two consecutive rounds. By introducing mirror residual encoding, it avoids judging behavioral status solely based on the number of connections and overall graph structure changes, thus distinguishing between normal changes in pedestrian density and abnormal dual-side interaction directions.

[0041] After the mirror-encoded state sequence is formed, the stable state structure corresponding to eight consecutive mirror partitions is extracted as the stable partition skeleton. Role migrations that occur within two mirror-encoded states and restore the original partition position are treated as short-term fluctuations. Each GraphScope state ownership encoding contention calls the six most recently formed stable behavior states, and compares the visual interaction description costs corresponding to the continuation of the current state, the return to a stable state, and the establishment of a new state. In the process of 92 stable behaviors reappearing, the conventional GraphScope was reclassified as a new graph segment 18 times, corresponding to a misclassification rate of 19.57%. This invention only failed to restore the original stable behavior state 4 times, corresponding to a misclassification rate reduced to 4.35%.

[0042] When the cost of establishing the visual interaction description corresponding to the new state is the lowest, the current mirror-encoded state is established as the new behavioral state, and the reversion blocking relationship is recorded. Local interaction structure stripping starts from the reversion blocking relationship and expands outwards by a maximum of 3 layers of direct connections. After removing one visual interaction connection each time, the cost of visual interaction description corresponding to stable state reversion and maintaining the new behavioral state is re-compared. When stable state reversion first achieves a cost advantage, the stripping range is stopped, and the stripped visual interaction connections are restored one by one for reverse structure reduction. When further reduction causes reversion encoding to lose its advantage, the minimum mismatch block is determined. In 126 abnormal behaviors, manually labeled abnormal interaction structures accounted for an average of 13.2% of the corresponding visual interaction connections, the average abnormal range determined by conventional GraphScope was 31.7%, and the minimum mismatch block determined by this invention accounted for an average of 14.6%, which is closer to the manually labeled range.

[0043] After the minimum mismatch block is determined, the background visual relationships outside the minimum mismatch block are restored to a stable behavioral state according to the stable partition skeleton. The minimum mismatch block continues to maintain its new behavioral state, and the scope of roles that need to re-participate in the GraphScope partition search is defined using the minimum mismatch block. Graph structure encoding changes and mirror residual encoding changes are then incorporated into the partition evaluation. The reverse adjustment is performed for a maximum of 8 rounds, and the adjustment stops when the minimum mismatch block and behavioral state remain unchanged for 3 consecutive rounds. In the test results, the average overlap between abnormal visual interaction structures and manually labeled abnormal ranges increased from 76.8% before the reverse adjustment to 84.9%, and the proportion of background visual relationships restored to a stable behavioral state increased from 91.3% to 97.6%, indicating that the reverse participation of the minimum mismatch block in the GraphScope partition search can reduce the continuous interference of local anomalies on normal background relationships.

[0044] Abnormal visual interaction states are ultimately reflected back to continuous monitoring video based on the identity and time location of the monitored target, and the abnormal behavior involving the target, the duration interval, and the range of visual interaction are output. Using manual review results as a benchmark, this invention is compared with anomaly identification methods based on target trajectory temporal features and conventional GraphScope dynamic graph anomaly identification methods. The anomaly identification method based on target trajectory temporal features extracts the position, speed, direction, and temporal change features of the continuous motion trajectory of the monitored target to determine anomalies in target behavior that deviates from normal motion patterns. The conventional GraphScope dynamic graph anomaly identification method uses GraphScope to partition and encode continuous dynamic graphs, identifies dynamic graph segment switching by changes in graph structure description cost, and identifies graph segments that significantly deviate from existing graph structure patterns as anomaly candidates. Specific comparison data is shown in Table 1. Table 1 Comparison of identification and location performance of different security monitoring anomaly detection methods

[0045] As shown in Table 1, the present invention achieves accuracy, recall, precision, and overall F1 score of 96.5%, 94.4%, 95.2%, and 94.8% respectively in abnormal behavior identification, all of which are higher than the trajectory temporal anomaly identification method and the conventional GraphScope method. Meanwhile, the false alarm rate for normal behavior is reduced from 7.6% in the conventional GraphScope method to 3.2%, and the false alarm rate for recurrence of stable behavior is reduced from 19.57% to 4.35%. This indicates that the competition between the stable partition skeleton and the state attribution encoding can effectively reduce false alarms caused by changes in the state of normal behavior. In terms of local anomaly identification… The local anomaly interaction localization accuracy of this invention reaches 91.3%, the average overlap ratio between the anomaly range and the manually labeled area reaches 84.9%, and the background visual relationship stability recovery ratio reaches 97.6%, which are higher than the conventional GraphScope method's 76.2%, 70.4%, and 89.7%, respectively. This indicates that minimum mismatch block identification and reverse partition adjustment can improve the localization accuracy of abnormal structures and reduce the impact of abnormal interactions on normal background relationships. Although the average processing time for a single graph state increases from 53.1 milliseconds to 61.7 milliseconds, it can still meet the dynamic graph processing requirements in continuous security monitoring scenarios.

[0046] 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 security monitoring anomaly identification method based on computer vision, characterized in that, Includes the following steps: Collect continuous monitoring video, perform visual positioning and tracking of the monitored targets, establish visual interaction relationships, and obtain a continuous visual interaction state sequence; The continuous visual interaction state sequence is converted into a visual relationship structure with a two-sided role correspondence, and a two-role visual relationship graph sequence is constructed. Based on the dual-role visual relationship graph sequence, the GraphScope dynamic graph mining algorithm is called to establish the role mirror correspondence, and the graph structure encoding change and mirror residual encoding change are jointly included in the partition search to generate the mirror encoding state sequence. For mirror-encoded state sequences, GraphScope state attribution encoding competition is performed. The behavioral state attribution is determined based on the visual interaction description cost, and the graph segment interpretation state is obtained. Based on the new behavioral state in the diagram interpretation state, perform local interaction structure stripping, and determine the minimum mismatch block state based on the condition that the recurrent encoding after stripping gains an advantage and loses its advantage after further reduction. Based on the minimum mismatch block state, restore the background visual relationship to a stable behavior state, retain the minimum mismatch block in the new behavior state, reverse the relationship between the role partition and the mirror residual encoding, and generate an abnormal visual interaction state. The abnormal visual interaction status is reflected back to the continuous monitoring video to determine the abnormal monitoring behavior and output the security monitoring anomaly identification result.

2. The security monitoring anomaly identification method based on computer vision according to claim 1, characterized in that, The specific steps to obtain the continuous visual interaction state sequence include: Visual positioning and continuous tracking of monitoring targets in continuous monitoring videos are performed according to the video time sequence. The spatial position and motion changes of the same monitoring target at continuous monitoring time are associated to obtain the target visual motion state sequence. Establish motion relationships between monitored targets in the target visual motion state sequence, continuously analyze the relative position changes, motion direction and relationship maintenance between targets, determine the visual interaction relationship with continuous motion relationship, and obtain the visual interaction relationship sequence; The visual interaction relationship sequence is state-sequenced according to the video time process, recording the establishment, maintenance, direction change and release of visual interaction relationships, and continuously organizing the interaction states corresponding to the same target identity to obtain a continuous visual interaction state sequence.

3. The security monitoring anomaly identification method based on computer vision according to claim 1, characterized in that, The specific steps involved in obtaining the dual-role visual relationship diagram sequence are as follows: Read the continuous visual interaction state sequence, determine the direction of action in the visual interaction relationship at each time step, establish a two-sided role mapping for the same monitoring target according to the starting position and pointing position of the visual interaction relationship, and maintain a fixed correspondence between the two-sided roles and the identity of the monitoring target to obtain the two-sided role state. For dual-role states, the visual interaction relationships in the continuous visual interaction state sequence are converted into role interaction relationships from the interaction initiator to the interaction recipient according to the direction of action. The role nodes of the interaction initiator and the interaction recipient are organized separately, and the corresponding role nodes are connected according to the role interaction relationship to construct a dual-role visual relationship diagram at each monitoring time. The dual-role visual relationship diagrams at each monitoring moment are organized continuously according to the time sequence of the continuous visual interaction state sequence. The identity indexes of the two sides corresponding to the same monitoring target are kept continuous in adjacent monitoring moments. The role interaction relationship between the two sides is updated according to the changes in the visual interaction state. The dual-role visual relationship diagrams at continuous monitoring moments are then linked sequentially to obtain the dual-role visual relationship diagram sequence.

4. The computer vision-based anomaly identification method for security monitoring according to claim 1, characterized in that, The specific steps to obtain the mirror-encoded state sequence include: Read the visual relationship graph sequence of the two roles, establish source role set and destination role set according to the interaction initiator side and the interaction receiver side respectively, call the GraphScope dynamic graph mining algorithm to perform initial partitioning on the source role set and destination role set, establish role mirror correspondence according to the identity correspondence of the same monitoring target in the two roles, establish mirror partition correspondence according to the role overlap relationship between the source partition and the destination partition with the same monitoring target identity, and obtain the role mirror partition status; The graph structure encoding changes of the two-sided partitions are determined based on the character mirror partition state, and the difference in graph structure encoding before and after the character partition adjustment is used as the structural encoding basis for GraphScope partition search to obtain the graph structure encoding state. Based on the graph structure coding state and the role mirror partition state, the visual interaction differences between the two sides of the same monitoring target under the corresponding mirror partition are compared to determine the mirror residual coding change. The graph structure coding change and the mirror residual coding change are jointly included in the GraphScope partition evaluation to obtain the joint partition coding state. For the alternating adjustment of the dual-side role partitions in the joint partition coding state, after the role partition on one side is adjusted, the mirror residual coding evaluation of the other side is updated according to the role mirror correspondence. The dual-side role partition adjustment is constrained by the reduction of joint coding cost to obtain the mirror coding state sequence.

5. The computer vision-based anomaly identification method for security monitoring according to claim 1, characterized in that, The specific process of obtaining the interpretation state of the diagram segment includes: Read the mirrored coding state sequence, compare the bilateral role partitions, role mirror correspondence, mirror partition correspondence and mirror residual coding relationship in the continuous mirrored coding state according to the monitoring time order, extract the partition structure that continuously maintains the stable mirror partition correspondence and does not have continuous role migration as the stable partition skeleton, and associate the stable partition skeleton with the corresponding behavior state to obtain the stable behavior coding state. For stable behavior encoding state and current mirror encoding state, partition inheritance is performed. Based on the role mirror correspondence, the stable partition skeleton is mapped to the current dual role visual relationship graph. For roles that maintain the same mirror partition correspondence, the partition in the stable partition skeleton is continued. For roles whose mirror partition correspondence has changed, partition inheritance is removed and they participate in GraphScope repartitioning to obtain the partition inheritance state. For partition inheritance states, GraphScope state attribution encoding competition is performed. Role partition search is constrained by stable partition skeleton, and behavior state attribution is determined based on the visual interaction description cost corresponding to different state attributions. When the visual interaction description cost corresponding to the establishment of a new state is the lowest, a new behavior state is established, and regression blocking relationships are recorded at the same time. The segment attribution of the current mirror-encoded state is determined based on the behavior state attribution, and the regression blocking relationship is retained when establishing a new behavior state to obtain the segment interpretation state.

6. The security monitoring anomaly identification method based on computer vision according to claim 1, characterized in that, Obtaining the minimum mismatch block state specifically includes: Read the diagram segment interpretation status, locate the roles and visual interaction connections involved in the reversion blocking relationship in the current dual-role visual relationship diagram, expand the local structure boundary along the directly connected visual interaction connections with the reversion blocking relationship as the starting position, and determine the local interaction structure stripping range. Regression-oriented stripping is performed on the local interaction structure stripping range. Starting from the visual interaction connection involved in the regression blocking relationship, the local interaction structure is gradually removed. The stopping condition is that the visual interaction description cost of the stable state regression is lower than the visual interaction description cost of maintaining the new behavior state for the first time, and a regressible stripping block is obtained. For reversible stripping blocks, reverse structural reduction is performed to gradually restore the visual interaction connection of the reversible stripping block boundary. The minimum mismatch block is determined based on the condition that the reversible encoding after stripping gains an advantage and the advantage is lost after further reduction. The minimum mismatch block is restored to its original structural position in the current dual-role visual relationship diagram, the association between the minimum mismatch block and the new behavioral state is maintained, and the structural boundary between the minimum mismatch block and the background visual relationship that can be explained by the stable behavioral state is recorded to obtain the minimum mismatch block state.

7. The computer vision-based anomaly identification method for security monitoring according to claim 1, characterized in that, The acquisition of the abnormal visual interaction state specifically includes: Read the state of the minimum mismatch block, maintain the structural position of the minimum mismatch block in the current dual-role visual relationship graph, and restore the visual interaction relationships outside the minimum mismatch block to the stable behavior state according to the stable partition skeleton; The minimum mismatch block is kept in the new behavioral state, and the original visual interaction connection between the minimum mismatch block and the background visual relationship is maintained. For the minimum mismatch block that remains in the new behavioral state, the adjustment range is limited by the roles involved in the minimum mismatch block and the visual interaction relationships directly connected with the background visual relationship. The graph structure encoding changes and mirror residual encoding changes are re-incorporated into the GraphScope partition evaluation, and the bilateral role partition and mirror residual encoding relationship are adjusted in reverse. Based on the adjusted bilateral role partitioning and mirror residual coding relationship, the minimum mismatch block and behavior state affiliation are re-determined, and the reverse adjustment is iteratively performed until the background visual relationship maintains a stable behavior state and the minimum mismatch block maintains the new behavior state, generating an abnormal visual interaction state.

8. The security monitoring anomaly identification method based on computer vision according to claim 1, characterized in that, The specific methods for obtaining the anomaly identification results from the security monitoring system include: Read the abnormal visual interaction status, and based on the identity and time location of the monitoring target corresponding to the smallest mismatch block, replay the continuous monitoring video to determine the range of abnormal visual interaction video; To recover the monitoring target and visual interaction relationship within the range of abnormal visual interaction videos, the continuous movement process of the monitoring target and the direction of visual interaction are correlated to determine abnormal monitoring behavior; Based on the time period of abnormal monitoring behavior, the monitoring target, and the visual interaction range, the identification information is organized and the security monitoring anomaly identification results are output.

9. A computer vision-based security monitoring anomaly identification system, comprising executing the computer vision-based security monitoring anomaly identification method according to any one of claims 1 to 8, characterized in that, include: The visual interaction construction module is used to acquire continuous monitoring video, perform visual positioning and tracking of monitoring targets, establish visual interaction relationships, and obtain a continuous visual interaction state sequence. The dual-role composition module is used to convert a continuous sequence of visual interaction states into a visual relationship structure with corresponding roles on both sides, and to construct a sequence of dual-role visual relationship graphs. The mirror coding partitioning module is used to call the GraphScope dynamic graph mining algorithm based on the dual-role visual relationship graph sequence to establish a role mirror correspondence, and incorporate graph structure coding changes and mirror residual coding changes into the partition search to generate a mirror coding state sequence. The state attribution competition module is used to perform GraphScope state attribution encoding competition for mirror-encoded state sequences, determine the behavior state attribution based on the visual interaction description cost, and obtain the graph segment interpretation state. The minimum mismatch identification module is used to perform local interactive structure stripping based on the new behavioral state in the graph segment interpretation state. The minimum mismatch block state is determined based on the condition that the recurrent encoding after stripping gains an advantage and loses an advantage after further reduction. The abnormal interaction generation module is used to restore the background visual relationship to a stable behavior state based on the minimum mismatch block state, retain the minimum mismatch block in the new behavior state, reverse the adjustment of the role partition and mirror residual coding relationship, and generate abnormal visual interaction states. The abnormal behavior output module is used to reflect abnormal visual interaction states back to continuous monitoring video, determine abnormal monitoring behavior, and output the security monitoring anomaly identification results.

Images