Gas field station abnormal behavior identification method fusing cross-camera behavior chain

Abstract

The present application relates to the technical field of intelligent video monitoring and artificial intelligence analysis, in particular to a gas station abnormal behavior recognition method fusing cross-camera behavior chains, comprising: collecting multi-path video streams, historical trajectory data, three-dimensional space topology graphs containing blind area nodes and operation regulation texts; extracting target appearance features and movement features, combining topology connectivity probability and dynamic tolerance to perform cross-camera joint matching, and generating identity labels; splicing trajectories according to time stamps, inserting blind area abnormal stay nodes in timeout cross-node segments, and forming space-time behavior paths; inputting path texts and video stream slices together with regulations into a visual language model to perform behavior intention reasoning; determining abnormal categories, calculating dynamic risk scores and triggering graded alarms; the present application organizes fragmented actions and blind area stays into inferable behavior chains, and realizes dangerous logic behavior recognition.

Description

Technical Field

[0001] This invention relates to the field of intelligent video surveillance and artificial intelligence analysis technology, specifically to a method for identifying abnormal behavior in gas stations by integrating cross-camera behavior chains. Background Technology

[0002] With the increasing demand for intelligent monitoring of gas stations, the identification of the activities of workers in the area covered by multiple cameras has gradually become an important part of safety management. Especially in scenarios such as nighttime maintenance, cross-regional inspections, and valve operation, how to accurately track the continuous behavior of workers and detect anomalies in a timely manner has become a technical problem that urgently needs to be solved in the current intelligent video analysis of gas stations.

[0003] Traditional abnormal behavior recognition technologies currently rely mainly on the following methods: target detection and recognition based on single camera images, cross-camera target re-identification based on appearance features, cross-regional trajectory stitching based on fixed time thresholds, and video anomaly determination based on single action segments. However, methods such as single-camera recognition, target re-identification based on appearance features, fixed-time threshold stitching, and single action segment judgment all have certain drawbacks. For example, single-camera recognition is difficult to reflect continuous behavior across regions, target re-identification based on appearance features is prone to mismatches under conditions of uniform clothing and occlusion of workers, fixed-time thresholds are difficult to adapt to different movement speeds and differences in site topology, and single action segment judgment is difficult to combine blind spot dwell information and operating procedures to make accurate judgments on illegal continuous operations. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for identifying abnormal behavior in gas stations by fusing cross-camera behavior chains. Specifically, the technical solution of this invention is as follows: A method for identifying abnormal behavior at gas stations by integrating cross-camera behavior chains includes: Step 1: Acquire multiple video streams, historical trajectory data, a 3D spatial topology map, and operation procedure text. The 3D spatial topology map includes blind spot nodes. Step 2: Extract the appearance and motion features of the target object from the multi-channel video streams, construct a directed topology graph based on the three-dimensional spatial topology graph, determine the dynamic tolerance based on the directed topology graph and the historical trajectory data, and generate the cross-camera identity identifier of the target object through multi-dimensional joint matching; Step 3: Based on the cross-camera identity identifier, stitch together the trajectory of the target object in the field of view corresponding to the multiple video streams and obtain the actual cross-node time. Based on the comparison result of the actual cross-node time and the dynamic tolerance, when the actual cross-node time is greater than the dynamic tolerance, insert the blind spot abnormal stay node corresponding to the blind spot node into the stitched trajectory to generate a spatiotemporal behavior path. Step 4: Based on the spatiotemporal behavior path, extract the multi-channel video streams corresponding to the time period to generate video stream slices. Input the spatiotemporal behavior path, the video stream slices, and the operation procedure text into a visual language model for analysis to generate behavioral intent reasoning results. Step 5: Determine abnormal behavior based on the behavioral intent reasoning results and calculate dynamic risk score, triggering the corresponding graded alarm instruction.

[0005] Preferably, in step two, the process of generating the cross-camera identity identifier of the target object through multi-dimensional joint matching includes: calculating the topological connectivity probability between nodes in the directed topology graph based on the historical trajectory data; and generating the cross-camera identity identifier through multi-dimensional joint matching using the appearance features, the topological connectivity probability, and the dynamic tolerance.

[0006] Preferably, the multidimensional joint matching performs feature fusion by comparing the similarity of the apparent features, the topological connectivity probability, and the time matching degree determined by comparing the actual cross-node time consumption with the dynamic tolerance, and outputs the best matching object as the cross-camera identity identifier.

[0007] Preferably, the abnormal dwell node in the blind zone includes a timestamp and spatial coordinate attributes, and the spatial coordinate attributes are generated by a preset coordinate mapping in the three-dimensional spatial topology map.

[0008] Preferably, in step four, the spatiotemporal behavior path is converted into a text description, and the text description is combined with the video stream slice to construct prompt words. The operation procedure text is converted into a structured knowledge graph, and the prompt words and the structured knowledge graph are input into the visual language model for analysis.

[0009] Preferably, the prompt word includes an instruction prefix, a contextual behavior description, and visual query content. The contextual behavior description consists of a textual description of the spatiotemporal behavior path, and the visual query content is used to guide the visual language model to recognize the operation actions in the video stream slice.

[0010] Preferably, the abnormal behavior includes bypassing to approach critical equipment, unauthorized continuous operation across regions, and abnormal violations in blind spots. The dynamic risk score is obtained by fusing the severity level of the abnormal behavior, the frequency of similar abnormalities within a preset time period, and the duration of the abnormal stay node in the blind spot.

[0011] Preferably, the triggering mechanism for the graded alarm command is as follows: When the dynamic risk score is greater than a preset first risk threshold, a system blocking control signal is generated for the controlled equipment in the gas station and a highest-level alarm data packet is output; when the dynamic risk score is greater than a preset second risk threshold and less than or equal to the preset first risk threshold, a manual review prompt message is generated and a medium-level alarm data packet is output; when the dynamic risk score is less than or equal to the preset second risk threshold, a regular log recording status flag is generated; wherein, the preset first risk threshold is greater than the preset second risk threshold.

[0012] Preferably, the method further includes a knowledge base update step: receiving misjudged labels from manual review feedback, extracting the spatiotemporal behavior path and the video stream slice corresponding to the misjudged labels as negative samples, and updating the parameters of the visual language model through a fine-tuning algorithm.

[0013] Preferably, the target is an operator wearing anti-static work clothes, and the operating procedure text includes a gas leak emergency handling process and valve operation specifications.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This solution is particularly suitable for multi-camera continuous operation scenarios such as gas stations and valve operations. It has the advantages of reducing the risk of mismatch between cameras when wearing anti-static work clothes and in blind spots, improving the ability to identify blind spot stops and restore cross-regional spatiotemporal behavior paths, enhancing the accuracy of judging abnormal behaviors such as illegal continuous operations, and realizing the output of hierarchical alarm commands. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0016] Figure 1 This is a flowchart of the present invention's method for identifying abnormal behavior in gas stations by integrating cross-camera behavior chains. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0018] like Figure 1 As shown, the method for identifying abnormal behavior at gas stations by fusing cross-camera behavior chains includes: Step 1: Acquire multiple video streams, historical trajectory data, a 3D spatial topology map, and operation procedure text. The 3D spatial topology map includes blind spot nodes. Step 2: Extract the appearance and motion features of the target object from the multi-channel video streams, construct a directed topology graph based on the three-dimensional spatial topology graph, determine the dynamic tolerance based on the directed topology graph and the historical trajectory data, and generate the cross-camera identity identifier of the target object through multi-dimensional joint matching; Step 3: Based on the cross-camera identity identifier, stitch together the trajectory of the target object in the field of view corresponding to the multiple video streams and obtain the actual cross-node time. Based on the comparison result of the actual cross-node time and the dynamic tolerance, when the actual cross-node time is greater than the dynamic tolerance, insert the corresponding blind spot abnormal stay node in the stitched trajectory to generate a spatiotemporal behavior path. Step 4: Based on the spatiotemporal behavior path, extract the multi-channel video streams corresponding to the time period to generate video stream slices. Input the spatiotemporal behavior path, the video stream slices, and the operation procedure text into a visual language model for analysis to generate behavioral intent reasoning results. Step 5: Determine abnormal behavior based on the behavioral intent reasoning results and calculate dynamic risk score, triggering the corresponding graded alarm instruction.

[0019] This embodiment provides a mechanism for identifying abnormal behavior in gas stations by integrating cross-camera behavior chains. Specifically, it is explained in the context of a nighttime inspection and maintenance scenario at the same gas station: the station is divided into a gas unloading area, a pressure regulating skid area, a main valve group area, and a booster pump area, each monitored by a camera. , , and Coverage; There are monitoring blind spots formed by pipe corridor obstruction between the main valve group area and the booster pump area; The objects to be identified are the operators who enter the station to perform inspections and valve operations. They generally wear anti-static work clothes and safety helmets of the same color. Relying solely on appearance characteristics can easily lead to misidentification. Therefore, it is necessary to link and process the cross-camera trajectory, blind spot stay, and operating procedures. To avoid ambiguity, physical monitoring blind spots are set up within the station. The corresponding blind spot node number in the 3D spatial topology map is... The system numbers the nodes in the blind zone. Based on the timeout result, insert an abnormal dwell event node. These three correspond to different expressions of the same blind zone area at three levels: region, topology node, and event node. Specifically, the system performs multi-source data acquisition; it simultaneously accesses multiple real-time video streams and reads trajectory data from at least one historical period, such as cross-regional passage records for all compliant inspection tasks within the last 30 days; concurrently, it imports a 3D spatial topology map of the site; this topology map is not a single planar set of nodes, but rather a spatial set of nodes composed of both camera-covered and blind-spot nodes; for example, nodes... Corresponding to the unloading area, node Corresponding pressure regulating skid area, node Corresponding main valve group area, node Corresponding to the booster pump area, node Corresponding blind spots in utility tunnels ;in, Coordinates are pre-stored in the 3D model, for example, coordinates (18.5, 7.2, 0.0); the operation procedure text includes rules such as inspection sequence, valve operation prerequisites, pressure relief waiting time, and whether a detector needs to be carried. Perform cross-camera target anchoring; the system extracts the appearance features and motion features of the target object from each frame of video; the appearance features can be a set of fixed-length vectors output by the feature extraction network, which can specifically adopt a pedestrian re-identification model based on a deep residual network. By pre-training on a large-scale human image dataset, it extracts fixed-length vector representations containing high-dimensional semantic features such as target clothing texture and body proportions; the motion features can include the direction of travel, average speed and gait rhythm when leaving the field of view. Then, construct a directed topological graph based on the three-dimensional spatial topological graph, for example, if there exists , , , Isodirectional edges; then, based on historical trajectory data, statistically analyze the travel time distribution and topological connectivity probability of each directional edge; with arrive For example, if historically 80% of compliance patterns are... The average time is 14 seconds, and the standard discrete range is ±4 seconds. Therefore, the connectivity probability of this path can be recorded as 0.8, and the baseline time window can be initially set to 10 to 18 seconds. When the target is handed over across cameras, the system performs multi-dimensional joint matching. For ease of explanation, assume that worker P is in Zhong left the main valve group area at 20:15:10, and the system was in Two candidates were found. and They entered the booster pump area at 20:15:23 and 20:15:41 respectively; if P and The apparent similarity between P and Q2 is 0.72, while the apparent similarity between P and Q2 is 0.81. Based solely on apparent similarity... The similarity score is higher; however, considering the topological relationship further, arrive The high-probability path is via Short-duration travel is not supported, and time intervals exceeding 31 seconds are not allowed. Furthermore, P's velocity before departure is 1.2 m / s, meaning the maximum dynamic tolerance at this speed is only 19 seconds. Therefore... The time matching score was high, while The weight was reduced due to timeout; The final system will still use P and Anchoring to the same identity avoids misassociation caused by uniform attire; then spatial behavior chain splicing is performed; the system sequentially splices the trajectories of the same identity in different cameras according to absolute timestamps; for example, identity identifier 17 in The trajectory in the image ends at 20:12:00. If the trajectory in the middle starts at 20:12:11, then the actual time spent crossing nodes is 11 seconds; if this time does not exceed the corresponding dynamic tolerance, then it is directly connected as a continuous path; if in another segment, ID-17 leaves at 20:15:10... Entered at 20:15:34 If the actual time taken to cross a node is 24 seconds, while the dynamic tolerance limit is 19 seconds, then the system determines that there is an invisible pause in that segment. In this case, the system according to and It must pass through This spatial topology allows for the insertion of blind zone abnormal stop nodes into the trajectory sequence. The start and end times of these nodes can be set to the portion exceeding 20:15:10 to 20:15:34. For example, with 19 seconds as the normal upper limit, the extra 5 seconds can be mapped to the duration of the abnormal stop. Their spatial coordinates are taken from the 3D topology map. The preset coordinates; further, the excess part in the aforementioned 20:15:10 to 20:15:34 specifically refers to the last segment of time after exceeding the normal passage limit, rather than recording the entire cross-regional period as abnormal stay; Therefore, in this example, the timestamps of the abnormal stop nodes are preferably marked as 20:15:29 to 20:15:34 to maintain consistency with the time attribution method used in the construction of subsequent blind zone nodes; thus, the resulting spatiotemporal behavior path is no longer just... , but Leaving Abnormal stay Arrival; execution of behavioral intent reasoning; the system extracts corresponding key segments from the original video stream based on the timestamps in the spatiotemporal behavioral path, for example, extracting the segment from 20:14:55 to 20:15:40. and The video stream slices were processed and the trajectory path was converted into a text description. For example, after the operator left the autonomous valve group area, he stayed in the blind area of ​​the pipe gallery for about 5 seconds, and then entered the booster pump area and performed a suspected valve contact action. Based on this, prompt words are constructed, which include both text descriptions and video stream slices corresponding to the path. These, along with the operating procedure text, are input into the visual language model as background constraints. The model not only identifies whether a single action is opening a valve or approaching equipment, but also comprehensively considers preceding trajectories, blind spot stops, and procedural context to determine whether it constitutes a violation chain. For example, the model may output that there is a failure to meet the pressure relief waiting time requirement before proceeding with subsequent actions to approach critical equipment, and that the action stops in a blind spot, indicating low compliance. Risk quantification and early warning output are then performed. The system first determines whether an anomaly is present based on the reasoning results, then classifies the anomaly type and calculates a dynamic risk score. For example, unauthorized continuous operations across regions are classified as high severity, detours to approach critical equipment are classified as medium to high severity, and abnormal violations in blind spots are classified as the highest severity. If the final score reaches a high level, a high-priority alarm is triggered, and an early warning data packet consisting of the identity identifier, behavior chain, video stream slice index, and risk conclusion is sent to the central control room. In addition, if communication with a certain camera is interrupted during a certain period, the system will still retain the existing topology and time constraints and mark that period as a visual missing segment. For the missing segment, only low-confidence stitching results are allowed to be generated based on adjacent cameras and topological reachability to avoid forcibly merging discontinuous trajectories directly. If the score difference between candidate matching objects is less than the preset safety boundary threshold, for example, the candidate object with the highest matching degree has a score of 0.61 and the candidate object with the second highest matching degree has a score of 0.60, then the identity chain will be set to a pending confirmation state and no high-level anomaly conclusion will be output for the time being. If a video stream slice is blurry or obscured, the system retains the text path reasoning results, but applies a confidence reduction to the risk score; if the operating procedure text is updated, the system will enable the new procedure version number and ensure that historical events can still be reviewed according to the old version of the procedure. During the night shift inspection and maintenance at this station, worker A should follow... Inspection of pressure regulating skid area Main valve assembly area observation pressure Waiting for pressure to be released The booster pump area confirmation process execution task; the system first identifies the same identifier 17 in C2 and C3, and discovers that it came from... arrive The transition time exceeds the dynamic tolerance and occurs in the blind spot of the utility tunnel. Insert abnormal dwell nodes; combine The video stream clip showing the person arriving at the scene immediately approaching the pump and making a gesture to rotate the valve, along with the textual restriction in the procedure that subsequent equipment cannot be started until the pressure relief waiting is completed, led the model to output that this behavior was an abnormal continuous operation; the system converted this result into a risk score and triggered a central control warning. The purpose of this step is to organize the fragmented actions and blind spots that were originally scattered across different cameras into a reasonable spatiotemporal behavior chain, thereby achieving accurate identification of dangerous logical behaviors of personnel. In this embodiment, the directed topology graph is constructed using the field of view coverage areas corresponding to multiple video streams as nodes. The travel time distribution is generated by statistical fitting of historical trajectory data. The movement features include speed features, and the dynamic tolerance is adaptively adjusted in real time according to the speed features.

[0020] This embodiment provides a mechanism for topological modeling of the camera coverage area and dynamic adjustment of the time window. Specifically, in the above-mentioned nighttime inspection and maintenance scenario at the station, if only a fixed number of seconds is used to determine whether two cameras belong to the same person, misjudgments are likely to occur in both fast-moving and paused observation states: when the fixed time window is less than or equal to the first set value, candidates arriving after normal slow movement or short pauses will be excluded; when the fixed time window is greater than or equal to the second set value, unrelated target personnel will be linked together. Therefore, the camera's field of view, historical traffic patterns, and real-time speed need to be incorporated into the judgment process; specifically, the nodes of the directed topology graph are directly abstracted from the actual area covered by the camera; for example... Covering the unloading area to form a node , Covering the pressure regulating skid area to form a node , Covering the main valve group area to form nodes , Covering the booster pump area to form nodes If there is a corridor or pipe gallery that is actually accessible to people between the coverage areas of the two cameras, then a directional edge is established; if it is not physically accessible, then no edge is established; the topology map constructed in this invention incorporates the site layout constraints, avoiding erroneous associations in areas of spatial physical obstruction; the passage time distribution is derived from historical trajectory statistical fitting. For ease of explanation, assume the system statistics include There were 100 valid passages, of which 60 fell between 7 and 10 seconds, 30 between 10 and 13 seconds, and 10 between 13 and 16 seconds. A time distribution centered around 10 seconds can be fitted. This time distribution can be interpreted as a statistical curve reflecting common passage times: the closer the time is to 10 seconds, the more consistent it is with historical compliant samples; as the time gradually deviates from 16 seconds, its support for normal passage decreases accordingly. This path, due to blind spots and equipment obstructions in the middle, may have a wider distribution, for example, mainly concentrated in 12 to 18 seconds; Dynamic tolerance refers to the normal time window boundary that the system allows a target object to move between two adjacent topological nodes when performing cross-camera trajectory stitching. In real-time applications, dynamic tolerance is not a fixed value, but is adaptively adjusted according to the moving speed before leaving the field. For example, if personnel are The average speed in the 3 seconds before departure is 0.6 meters per second, and the system can... The dynamic tolerance is set to 15 to 22 seconds; if another person's speed reaches 1.4 m / s before leaving the field, the time window for the same path can be tightened to 9 to 16 seconds; the speed here is not the instantaneous value of a single frame, but the smoothed average of several consecutive frames, in order to reduce the error caused by jitter. The principle is that for the same road, fast movement and slow travel usually correspond to different reasonable travel times. Furthermore, real-time adaptive adjustment does not determine the time window independently of historical distribution, but uses the common interval given by historical travel time distribution as a benchmark, and then combines the current target's movement speed before leaving the field to tighten or loosen the benchmark interval. Specifically, the system extracts historical passage time distribution to obtain a baseline time interval, and dynamically adjusts the upper and lower limits of tolerance according to the current target's real-time movement speed using the following formula: In the formula: The average speed of the historical trajectory of this path, i.e., the standard reference speed, has the dimension of . ; The continuous smooth movement speed of the current target before it leaves the field, with dimensions of ; The time scaling factor is determined based on the velocity ratio and is dimensionless. and These are the lower and upper limits of the baseline time interval extracted from historical data, respectively, with dimensions of... ; and These are the lower and upper limits of the dynamic tolerance output after the system's adaptive adjustment, respectively, with dimensions of... .

[0021] In other words, the historical distribution reflects the baseline travel time of the path, while the real-time speed reflects the deviation of the current target's actual movement rate from the standard reference speed, thus avoiding over-adjustment of the upper and lower limits of dynamic tolerance based solely on instantaneous speed. In addition, if a person pauses briefly before leaving, causing the speed measurement to approach zero, the system does not directly expand the dynamic tolerance to the maximum. Instead, it uses preset upper and lower limits to limit the threshold. For example, the lower limit is still no less than the 10th percentile of the historical distribution, and the upper limit is still no higher than the 90th percentile. If there are insufficient historical samples, such as only 3 valid trajectories in a newly established area, the initial distribution shared with similar areas is used first, and then gradually replaced after accumulating samples. If two areas are topologically adjacent but have been almost untouched historically, the connectivity probability is automatically reduced, and they will not be prioritized for matching even if the time is close. Furthermore, if an unstable travel state such as looking back, turning around in place, or pausing to observe is detected before leaving the field, the system can mark the time period as unstable speed reference. In this case, instead of directly using the instantaneous state to expand the time window, the system will roll back to use a smooth speed over a longer period of time or directly use the historical reference time window to avoid misinterpreting local motion noise as changes in cross-regional traffic capacity. When operator A moves from the pressure regulating skid area to the main valve assembly area, the system reads his / her position... The average moving speed in the game is 1.1 meters per second, corresponding to... The dynamic tolerance was set to 8 to 12 seconds; the actual arrival time was 9 seconds, therefore it was considered a normal transition; however, before it moved from the main valve group area to the booster pump area, the system detected a noticeable pause and inspection action before leaving, with the average speed dropping to 0.7 m / s, so it was... The dynamic tolerance has been increased to 14 to 21 seconds; if the final time reaches 24 seconds, it will still fall within the abnormal range and will trigger blind zone dwell reasoning. The purpose of this mechanism is to couple the accessibility of the site space with the real-time movement status of personnel, so as to achieve cross-camera timing judgment that is more in line with the actual operation rhythm and reduce false connections and missed connections caused by fixed thresholds. In this embodiment, the multidimensional joint matching performs feature fusion by comparing the similarity of appearance features, the topological connectivity probability, and the time matching degree determined by comparing the actual cross-node time consumption with the dynamic tolerance, and outputs the best matching object as the cross-camera identity identifier.

[0022] This embodiment provides a weighted fusion mechanism for multi-dimensional joint matching. Specifically, based on the above, if cross-field target association is performed solely based on appearance feature similarity, in gas stations, the same target will frequently be assigned multiple identity identifiers or multiple targets will share the same identity identifier due to identical tooling, nighttime illumination below the preset illuminance threshold, and safety helmet obstruction. Therefore, it is necessary to incorporate appearance similarity, topological reachability, and temporal matching into a unified scoring system. Specifically, the system extracts features for each pair of candidate cross-camera targets and calculates the final fusion score using the following formula: In the formula: The final multidimensional joint matching score for the candidate targets is dimensionless. The apparent similarity score is calculated based on the distance between the feature vectors of the targets. It is dimensionless and its value ranges from [value missing]. ; The topology connectivity probability score reflects the probability of the current jump path being connected in the historical topology; it is dimensionless and its value ranges from [value missing]. ; The time matching score reflects the degree of fit between the actual time spent across nodes and the dynamic tolerance; it is dimensionless and its value range is [value range missing]. ; The preset weight coefficients for the above three dimensions are respectively, satisfying... .

[0023] For ease of understanding, let's assume that... The departing target P has two candidate successors. and P and The apparent score is 0.74, the topological connectivity probability is 0.90, and the time matching score is 0.88; P and The apparent score is 0.81, the topological connectivity probability is 0.35, and the time matching score is 0.20; calculated according to the above weights, The overall score is 0.74×0.5+0.90×0.2+0.88×0.3=0.814; The overall score is 0.81×0.5+0.35×0.2+0.20×0.3=0.535; although While exhibiting higher similarity in apparent features, its routes and times do not conform to the preset topological connectivity rules and dynamic tolerance constraints. Therefore, it should ultimately be selected. Generate cross-camera identity identifiers; Furthermore, the time matching score can be calculated using an interval proximity method; if the actual cross-node time falls within the dynamic tolerance, a score higher than the preset matching threshold is assigned; the closer to the center value of the time window, the higher the score; if it exceeds the time window, the score decreases according to the preset decay function; this can distinguish between two independent states: the actual cross-node time of the target object is within the dynamic tolerance but deviates from the statistical optimal value, and the actual cross-node time exceeds the upper limit of the dynamic tolerance; the topology connectivity probability score comes from historical traffic statistics. If a certain jump almost never occurs, even if it is on the edge of the dynamic tolerance in time, it will not become a high-priority candidate; In a further fault-tolerance mechanism, if the combined scores of multiple cross-camera matching objects are all below the minimum acceptance threshold, for example, all less than 0.45, the system will not force the assignment of identities, but will interrupt the current chain and wait for subsequent more definitive observations to avoid the spread of misassociations to the entire behavior chain; if the difference between the scores of the first and second place is less than the safety boundary, for example, less than 0.03, the segment will be marked as a low-discrimination handover and sent to the manual review queue or the subsequent multi-frame re-scoring mechanism; if a certain score is missing, for example, due to occlusion preventing stable extraction of appearance features, the weight of the topology and time dimensions can be temporarily increased, but the final confidence level will be lowered at the same time. After worker A left the main valve group area at 20:15:10, the system detected two workers wearing the same work clothes in the booster pump area. Judging from the images alone, the later worker was closer to worker A. However, considering the probability of a short connection between the main valve group area and the booster pump area via the pipe gallery, and the tight time window corresponding to worker A's faster walking speed before leaving the site, the system calculated that the overall score of the worker who appeared first was higher. Therefore, the identity chain was kept continuous, and another irrelevant person was not mistakenly included in the abnormal behavior path. The purpose of this mechanism is to introduce the spatial knowledge and temporal patterns of the site into the identity matching process through quantifiable multi-factor fusion, so as to achieve more robust cross-camera identity anchoring. In this embodiment, the abnormal dwell node in the blind zone includes a timestamp and spatial coordinate attributes. The timestamp is determined by the time difference, and the spatial coordinate attributes are generated by a preset coordinate mapping in the three-dimensional spatial topology map.

[0024] This embodiment provides a mechanism for constructing abnormal dwelling nodes in blind spots. Specifically, in the aforementioned path stitching, if the disappearance from one shot to the appearance of another shot is considered as a normal cross-camera transition, then all dwelling, turning back, contact with equipment, or suspicious destructive actions hidden in the blind spot will be reduced to a single spatial displacement feature, and subsequent intent analysis will lose important evidence. Therefore, this embodiment explicitly abstracts the abnormal time period in the blind spot into a computable node. Specifically, when the system detects that the actual time spent across nodes exceeds the dynamic tolerance, the excess time is calculated; for example... The tolerance limit is 19 seconds. The actual time spent across nodes is 24 seconds, so the excess time is 5 seconds. The system can register this excess as an abnormal dwell time in the blind zone and insert an intermediate node in the behavior chain. This node has timestamp and spatial coordinate attributes. The timestamp attribute can represent a dwell time interval or a combination of the start time and duration of the anomaly. For example, if the departure time is 20:15:10, and the 19 seconds within this interval is within the normal time frame for crossing zones, then the start time of the anomaly can be marked as 20:15:29, and the end time as 20:15:34. The spatial coordinate attribute is directly mapped to the 3D topology map. The preset coordinates of the node, for example (18.5, 7.2, 0.0); It should be noted that the marking method for the period from 20:15:29 to 20:15:34 is essentially to structurally register the time difference exceeding the normal passage limit as an abnormal stay timestamp for subsequent behavioral chain reasoning; its meaning is that this 5-second abnormal time consumption belongs to This corresponds to the blind spot, not that the system has accurately captured the person's actual microscopic positional changes within the blind spot every second; that is, the node construction in this embodiment emphasizes calculable attribution of abnormal time consumption, rather than fabricating a reconstruction of trajectory details within the invisible blind spot; after this construction, the behavior path changes from ordinary Expand to During subsequent text conversion, you can... The interpretation is that there was an abnormal stay of 5 seconds in the blind zone of the pipe gallery between the main valve group area and the booster pump area, rather than a generalized description of increased passage time. This is especially important for visual language models to understand procedural violations, because the model can use this to associate the blind zone stay with the approach of critical equipment. Furthermore, if the same blind zone has multiple candidate coordinate points, such as a blind spot in the pipe gallery with a valve box on the left and a blind spot in the maintenance passage on the right, the system can select the most likely preset coordinates based on the departure and entry directions; for example, if personnel are traveling from... Exit from the right boundary and from If the entry point is near the pump body inlet, the coordinates of the blind spot in the right maintenance channel will be mapped first; if the directional information is insufficient, the coordinates of the geometric center of the blind spot will be used, while retaining the coordinate uncertainty marker. Furthermore, when a cross-regional process may theoretically cover multiple consecutive blind zones, the system does not simply stack all timeouts onto a single coordinate point. Instead, it prioritizes selecting a primary blind zone node based on the shortest path in the topology and consistency of the entry and exit directions. Only when there is a clear relationship of multiple blind zones passing through each other in the topology will the system be split into multiple blind zone nodes in sequence to keep the spatiotemporal behavior path consistent with the actual site structure. In the further fault tolerance mechanism, if the actual time spent crossing nodes is only slightly higher than the tolerance limit, for example, less than 1 second, then an abnormal stop node will not be generated immediately. Instead, it will be recorded as an edge timeout. It will be upgraded only after it occurs multiple times in a row, so as to avoid false alarms caused by time synchronization jitter. If the coordinates of the blind zone node are not preset in the 3D topology map, the system will not arbitrarily fabricate coordinates. Instead, it will temporarily point the spatial attribute to the number of the blind zone area and prompt for subsequent map completion. If a cross-zone process may pass through multiple blind zone nodes, it can be assigned to a single node according to the shortest path and direction consistency, or split into multiple nodes according to the time ratio. During the process of worker A moving from the main valve assembly area to the booster pump area, the system detected a total time of 24 seconds, while the normal upper limit based on his walking speed was 19 seconds. Therefore, it automatically inserted a checkpoint in the pipe gallery blind spot. The node, with timestamps from 20:15:29 to 20:15:34, is mapped to a preset location in the pipe gallery's obstruction zone. This node is written into the spatiotemporal behavior path and becomes an explicit input for subsequent inference regarding whether unauthorized contact or stay occurred in the blind zone. The purpose of this mechanism is to incorporate previously invisible blind zone behaviors into the analysis chain in the form of structured nodes, thereby achieving explicit modeling of hidden risks. In this embodiment, the visual language model aligns the text description of the spatiotemporal behavior path with the visual features of the video stream slices through a multimodal attention mechanism, and outputs a behavior intent reasoning result that includes compliance determination.

[0025] This embodiment provides a mechanism for using operational procedure knowledge graphs for multimodal compliance reasoning. Specifically, in the aforementioned method, even if a complete behavioral chain has been obtained, if anomalies are judged solely based on the video actions themselves, normal inspection stops and continuous violations may still be misclassified. This is because the risks in industrial settings do not depend solely on the apparent visual characteristics of actions, but also on the sequence of actions, regional constraints, and preconditions. Therefore, this embodiment introduces operational procedure knowledge graphs and multimodal alignment analysis. Specifically, the system pre-organizes the operation procedure text into a structured knowledge graph. Regions, equipment, actions, prerequisites, prohibitions, and waiting times can be represented as different entities and relationships within the graph. For example, the graph might contain relationships such as: main valve group area - execute action - pressure confirmation; pressure confirmation - prerequisite - pressure relief waiting; pressure relief waiting - duration not less than - 30 seconds; incomplete pressure relief waiting - prohibition - approaching booster pump area valves. In this way, the procedure is no longer unstructured text, but a structured rule network that can be indexed by the model. When spatiotemporal behavior paths and corresponding video stream slices are input into the model, the multimodal attention mechanism aligns key elements in the text with target segments in the visual representation. Specifically, the visual language model extracts the spatiotemporal feature maps and text word vectors of knowledge graph entities from video stream slices through a visual encoder and a text encoder, respectively. It uses a cross-attention layer to calculate the similarity weights between text word vectors and spatiotemporal feature maps in different spatial and temporal dimensions, thereby accurately mapping text instructions to the corresponding video frames and spatial regions, achieving feature alignment between key elements and target segments. For example, the text description mentions leaving the main valve group area, pausing in a blind zone, and then entering the booster pump area. The visual segment detects independent action features such as right-hand rotation of the valve handle, absence of detection instruments, and immediate approach to the pump body upon entry. During alignment, the model connects signals such as the booster pump area's valve handle not completing its waiting period with prohibition relationships in the knowledge graph. The final output includes not only the identification result of the specific action category but also a judgment on whether the action conforms to the procedure. To avoid conflict with the node numbering in the aforementioned 3D spatial topology graph... , , , To avoid confusion, in the following deduction, the three timing events in the spatiotemporal behavior path—leaving the main valve group area, staying in the blind zone for 5 seconds, and contacting the booster pump area valve—will be sequentially recorded as timing events. Sequence events and timing events The three sets of action evidence extracted from the video stream slices—observation of the pressure gauge, blind spot with no video image, and rotating valve handle—were sequentially recorded as action evidence. Action evidence and action evidence ; Based on this, a simplified deduction can be made: after model alignment, two key chains are obtained, one of which is a time-series event. With action evidence The first is to jointly support the completion of the preliminary checks; the second is to address the timing events. With action evidence They jointly supported the implementation of subsequent operations; however, the graph requires time-series events. The intermediate condition of waiting 30 seconds must then be met, but the current path only shows a 5-second blind spot, therefore the output procedure is not met, resulting in low compliance; among other things, action evidence... The absence of video images in the blind spots indicates a lack of direct visual observation evidence for that period. The corresponding risk is mainly based on the structured blind spot dwelling events registered in the path, rather than mistaking invisible periods for specific hand movements that have been observed. In the further fault tolerance mechanism, if the video stream slice cannot clearly identify the specific hand action, the model can still give a conservative conclusion based on the path text and regional context, such as outputting that there is a procedural risk but insufficient action evidence, and lowering the confidence evaluation value of the conclusion to below the preset confidence threshold; if there are multiple conflicting versions of the procedural text, the effective version associated with the task's corresponding timestamp shall prevail; if the knowledge graph has not yet included a certain type of new device rule, the model will only perform basic action recognition for that segment and will not directly output a strong compliance conclusion to avoid over-inference; In the event chain of operator A, the system encodes the following actions into text: leaving the main valve group area at 20:15:10, abnormal stop in the blind zone from 20:15:29 to 20:15:34, and entering the booster pump area and performing a rotary valve action at 20:15:34. Simultaneously, it extracts... and The key video stream slices; the knowledge graph clearly stipulates that after completing the operation of the main valve group area, a pressure relief waiting period of no less than 30 seconds is required, and no subsequent operation shall be carried out on the equipment in the booster pump area before the pressure relief is confirmed; based on this, the model outputs that there is a continuous operation violation, and the compliance is judged as failing. The purpose of this mechanism is to transform industrial procedures from background information into structured constraints that can be used for reasoning, thereby enabling behavioral intent analysis for site safety, rather than remaining at the level of single-action recognition. In this embodiment, the prompt word includes an instruction prefix, a contextual behavior description, and a visual query content. The contextual behavior description consists of a textual description of the spatiotemporal behavior path, and the visual query content is used to guide the visual language model to recognize specific operation actions in the video stream slice.

[0026] This embodiment provides a prompt word construction mechanism adapted to industrial safety scenarios. Specifically, based on the above, if only video stream slices and a piece of original text are directly fed into the model, the model may focus on irrelevant details, such as the color of personnel's clothing, background vehicles, or ordinary walking movements, while ignoring core detection targets such as whether valves are touched or whether detectors are carried. Therefore, this embodiment uses layered prompt words to clearly guide the model's attention to the information required for procedure review. Specifically, the prompt words contain at least three parts; the first part is the instruction prefix, which is used to limit the model task type, such as requiring it to combine the site regulations to conduct compliance review of the target personnel's cross-regional behavior and output the anomaly category, basis and confidence level; the second part is the context behavior description, which is transcribed from the spatiotemporal behavior path, and does not just list the location points, but retains the time sequence, stay information and regional relationship. For example, it can be described as follows: The target personnel briefly stopped in the pressure regulating skid area at 20:12, left the main valve group area at 20:15, and after abnormally stopping for about 5 seconds in the pipe gallery blind area, entered the booster pump area at 20:15:34; The third part is the visual query content, which is used to clearly define the key points of the actions that the model should prioritize for identification. For example, please focus on judging whether there are actions such as rotating valve handles, operating equipment panels, carrying detectors, or touching the pump body guardrail with both hands. A simplified example can illustrate the difference; if the input is simply "Analyze this video," the model might only return that a worker is moving near the equipment; however, with the structured cue words mentioned above, the model is more likely to output that the worker rotated the valve handle with his right hand within 3 seconds of entering the booster pump area, without any waiting behavior or portable detection instruments being observed; it can be seen that the visual query content in the cue words plays a role in task focus, while the contextual behavioral description provides the logical background before and after the action occurs. Furthermore, prompts can also embed blind spot information; for example, explicitly stating in the context that there is a non-visual blind spot in the middle, so that when the model sees subsequent device contact actions, it will not regard the action as an isolated event, but will actively connect it to the potential preceding behavior in the blind spot; this can improve the ability to analyze complex violation chains. In addition, if the number of nodes or the time span of the spatiotemporal behavior path exceeds the preset segment threshold, the system will not write the entire path into the prompt words at once. Instead, it will extract the time window most relevant to the current action to be reviewed, such as the first 30 seconds to the last 10 seconds, to avoid the distraction caused by excessively long context. If the video resolution is lower than the preset resolution lower limit required for visual query features, the system will automatically simplify the query items, such as degrading from judging the valve rotation angle to judging whether there is an obvious valve operation action. If there is a discrepancy between the text description and the video timestamp, the prompt words will be generated after recalibration with a unified absolute timestamp. In the review of worker A's behavior, the system-generated instruction prefix is ​​"Please determine whether the target person has engaged in continuous unauthorized operations in accordance with the gas station safety regulations." The contextual behavior description is that after completing the stop in the main valve group area, the person abnormally stopped for 5 seconds in the pipe gallery blind spot before entering the booster pump area. The visual query content focuses on identifying whether there is any approach to critical equipment, manual rotation of valves, or missing detection instruments. Based on this, the model provides a more focused compliance analysis conclusion than ordinary video question and answer. The purpose of this mechanism is to improve the model's attention to key industrial actions and logical relationships through a controllable prompt structure, thereby achieving a more stable output for procedure review. In this embodiment, abnormal behaviors include bypassing to approach critical equipment, unauthorized continuous operation across regions, and abnormal violations in blind spots. The dynamic risk score is obtained by fusing the preset quantitative value corresponding to the severity level of the abnormal behavior, the frequency of occurrence, and the duration of abnormal stay at the blind spot node. The severity level is determined and output by the visual language model combined with the operation procedure text, and the frequency of occurrence is the number of times within a preset time period.

[0027] This embodiment provides an abnormal behavior classification and dynamic risk scoring mechanism. Specifically, after the aforementioned intent reasoning, if the system only outputs binary conclusions such as abnormal or normal, the central control personnel still find it difficult to distinguish whether it is a minor detour, repeated violation, or a high-risk continuous operation that is very likely to induce a major safety accident. Therefore, it is necessary to further subdivide the abnormal types and compress multidimensional risk factors into a single score that is convenient for early warning decision-making. Specifically, abnormal behavior can be divided into at least three categories: the first category is detouring to approach critical equipment, which means that the target personnel enter the equipment area that they should not approach by following the normal route; the second category is unauthorized continuous operation across areas, which means that the operation of equipment in a subsequent area is carried out immediately before the operation in a previous area has met the requirements of the procedure; the third category is abnormal violation of blind spot, which means that after staying in a blind spot for an abnormally long time, there are signs such as abnormal approach to equipment, climbing of guardrails or suspicious tool movements. After classifying abnormal behaviors, the system constructs a dynamic risk score based on multidimensional risk factors, and the calculation formula is as follows: In the formula: The output is a dynamic risk score, which is dimensionless. This is a preset quantitative value corresponding to the severity level of this type of abnormal behavior; it is dimensionless. The normalized value of the frequency of similar anomalies occurring under the same identity within a preset time period is dimensionless. The duration of the abnormal dwell time node in the blind zone during this event, with dimensions of ; The maximum reference abnormal dwell time preset for the system is used to normalize the duration item, with the dimension being... ; The weighting coefficients for each risk dimension satisfy the following conditions: This formula ensures that any extreme mutation of a single factor can be effectively mapped to the risk score.

[0028] For example, if an incident is classified as high severity, with a quantification value of 0.9; if two similar violations have occurred in the last two hours, the normalized frequency is 0.4; and if the abnormal blind spot lasted for 12 seconds, which is normalized to 0.4 based on a 30-second timeframe, then the total score is... The advantage of this calculation method is that it will not immediately trigger the highest risk due to a brief edge anomaly, but it will not ignore repeated medium-level violations. If the same person repeatedly detours to approach critical equipment, even if the single stay is short, the cumulative frequency will significantly increase the total risk score. If a blind spot stay continues to increase significantly, even if the frequency was not high before, the risk score will increase exponentially. Furthermore, historical events with the same anomaly category as the current one should be prioritized for frequency statistics. If business needs require different anomaly categories to be included in the same risk chain, the system should first preset the association category mapping rules. For example, marking detours to approach key equipment and unauthorized continuous cross-regional operations as the same continuous violation risk cluster should be used before merging and counting them, rather than arbitrarily accumulating the number of different categories without any rules. This can ensure that the frequency item is consistent with the anomaly classification logic and avoid the drift of the scoring criteria. In the further fault tolerance mechanism, if a certain type of anomaly does not have a blind zone node, such as unauthorized continuous operation within a purely visible area, the blind zone duration term can be set to 0 without affecting the calculation of the other two terms; if there are no historical anomalies within a preset time period, the frequency term is set to 0 to avoid program errors caused by division by zero during the calculation process; a fixed reference upper limit is used as the denominator during normalization; if the model's confidence in the severity level determination is too low, the system can backtrack the severity level quantization value to the adjacent lower level to reduce false alarms; if the same event hits multiple anomaly categories at the same time, the highest severity level can be prioritized, or the upper limit of the union can be taken after scoring them separately. In this nighttime incident involving worker A, the visual language model determined, based on the procedural text, that it constituted unauthorized continuous operation across regions, with a high severity level and a corresponding quantitative value of 0.9. The system found that he had performed the same type of unauthorized continuous operation once in the past 2 hours, so the frequency normalization result was recorded as 0.3. If the business side has pre-defined the detour approaching the booster pump area as a preceding related event of this type of continuous violation, it can also be included in the statistics according to the association category mapping rules. The abnormal stay in the blind spot lasted 5 seconds, corresponding to a duration normalization value of approximately 0.17. After comprehensive analysis, a medium-to-high risk score was obtained, which is sufficient to trigger further action. The purpose of this mechanism is to uniformly map abnormal evidence from different dimensions into comparable and thresholdable risk scores, thereby achieving more granular safety management. In this embodiment, the triggering mechanism for the graded alarm command is as follows: When the dynamic risk score is greater than the preset first risk threshold, a system blocking control signal is generated for the controlled equipment in the gas station and the highest level alarm data packet is output. When the dynamic risk score is greater than the preset second risk threshold and less than or equal to the preset first risk threshold, a manual review prompt message is generated and an intermediate alarm data packet is output. When the dynamic risk score is less than or equal to the preset second risk threshold, a regular log record status flag is generated, wherein the preset first risk threshold is greater than the preset second risk threshold.

[0029] This embodiment provides a graded alarm triggering mechanism linked to risk scoring. Specifically, after the aforementioned risk quantification, if all anomalies are alarmed using the same method, the following problem will occur: minor suspicious behaviors frequently interfere with on-site operations and cannot be addressed within the preset safety response timeframe. Therefore, it is necessary to differentiate the output method according to the risk level. Specifically, the system sets at least two thresholds: the preset first risk threshold corresponds to the high-risk range, and the preset second risk threshold corresponds to the range of concern, with the former being greater than the latter. For ease of explanation, the preset first risk threshold can be set to 0.75, and the preset second risk threshold can be set to 0.45. When the dynamic risk score is greater than 0.75, the system generates the highest-level alarm data packet and outputs a blocking control signal to the relevant control unit. The blocking method can be selected according to the site safety design as one or more combinations of access control locking, remote equipment operation permission freezing, audible and visual alarm activation, or forced pop-up window on the central control screen. When the score is between 0.45 and 0.75, the system does not immediately execute forced blocking, but pushes a manual review prompt and outputs a medium-level alarm data packet for the duty officer to quickly confirm. When the score is not higher than 0.45, the event is only logged and subsequently statistically analyzed, without interfering with on-site operations. The alarm data packet may include identification, time, location chain, anomaly category, risk score, keyframe index, and procedure basis summary. In this way, central control personnel can directly view the specific reasons and logical basis for the alarm triggering without having to search through massive amounts of video footage. For the highest level alarm, the coordinates and duration of the abnormal node in the blind spot can also be attached, making it easier for security personnel or inspection personnel to quickly go to the corresponding area to handle the situation. In the further fault tolerance mechanism, if the station is currently in a special operating condition, such as emergency repair mode, the system can call different threshold templates to prevent normal emergency operations from being blocked as violations; if the blocking control signal fails to be sent, the system should retain the failure status and escalate the notification to human intervention, instead of assuming the handling is completed; if no one confirms the intermediate alarm within a preset time, the priority can be automatically increased according to the timeout escalation strategy; if the same identity triggers multiple low-level events in a short period of time, the subsequent events can also be escalated to the human review level through the cumulative strategy. In the case of worker A, if the risk score calculated by the system is 0.82, it exceeds the preset first risk threshold. The system immediately sends the highest-level alarm data packet to the central control room and freezes the remote operation authorization under the corresponding work ticket. At the same time, the behavioral chain of leaving the main valve group area, staying in the pipe gallery blind area, and operating the valve in the booster pump area is highlighted on the monitoring display terminal. If another incident is only a suspicious detour with a risk score of 0.58, only a manual review prompt will pop up. The duty officer will review the video stream slices and decide whether to escalate the action. The purpose of this mechanism is to take different intervention measures according to the strength of the risk, so as to achieve a balance between alarm effectiveness and on-site availability. In this embodiment, a knowledge base update step is also included, which involves receiving misjudged labels from manual review feedback, extracting the spatiotemporal behavior path and video stream slices corresponding to the labels as negative samples, and updating the parameters of the visual language model through a fine-tuning algorithm.

[0030] This embodiment provides a knowledge base update mechanism based on manual review and feedback. Specifically, during long-term operation in industrial sites, changes in site layout, adjustments to work processes, and camera angle drift may all lead to new types of misjudgments in the model. If the system always uses a fixed model, the judgment rules that were effective in the early stages may gradually deviate from reality under subsequent working conditions. Therefore, this embodiment introduces a closed-loop update process. Specifically, when a medium or high alarm enters the manual review stage, the duty officer can label the event with tags such as false alarm, missed alarm, supplementary alarm, or inaccurate classification. After receiving the false alarm label, the system backtracks to extract the spatiotemporal behavior path, video stream slice, procedure version number, and the reasoning conclusion output at that time corresponding to the event, forming a trainable sample. If the event is ultimately confirmed to be a false alarm, it can be added to the update set as a negative sample. The negative sample here is not just a single frame image, but a joint sample of path text, video stream slice, and manual conclusion, which can better correct the model's deviation at the logical reasoning level. Furthermore, to maintain consistency with the update method for negative samples in this document, tags that are missed or inaccurately classified are first stored as verification samples or samples to be processed when they first enter the system, and are not directly merged into the negative sample set of this round; the system performs initial screening of tags by establishing a temporary sample buffer, and only when a tag can be clearly identified as a false alarm after review and merging is it transferred to the negative sample pool for subsequent fine-tuning according to the false alarm sample caliber; For example, in a certain emergency repair situation, worker A entered the booster pump area before completing the normal waiting time, but his work order indicated that the emergency procedure had been activated, so the behavior was actually compliant; if the model reports a high risk because it does not fully understand the emergency mode, the event can be marked as a false alarm after manual review; the system writes the path text, video stream slices and background information on the allowed shortening of waiting time under emergency conditions into the updated sample for subsequent fine-tuning. In this way, the model will gradually achieve adaptive differentiation between normal and special working conditions. During fine-tuning, the system can adopt a small-batch incremental update strategy and a parameter-efficient fine-tuning algorithm. Under the premise of freezing the parameters of the main network of the visual language model, only the weights of the newly added bypass network or soft cue vectors are updated with gradients to avoid overfitting of the model due to the number of samples in a single batch being lower than the preset batch threshold. The updated model version needs to be verified in parallel with the original version for a period of time. After confirming that the false positive rate has decreased and the false negative rate has not increased significantly, the online version can be replaced. The knowledge base can also be supplemented with new rule nodes or new abnormal example descriptions in a synchronous manner. In the further fault tolerance mechanism, if human feedback is contradictory, such as different reviewers giving different labels to the same event, the system will not directly include it in training, but will instead enter the review arbitration process; if the number of negative samples is lower than the preset minimum sample size for a single fine-tuning, it will be cached in the sample pool and updated uniformly after the minimum batch size is reached; if the performance on the validation set deteriorates after the update, it will be rolled back to the previous stable version; if some misjudgments are not due to model problems, but rather to time synchronization errors or camera occlusion, the front-end data link should be corrected first to avoid executing invalid parameter updates. After one month of operation at this site, the system misjudged a cross-regional coordinated operation with an authorized work ticket as unauthorized continuous operation. The shift operator marked the event as a false alarm in the review interface and added that the shift was in the leakage emergency response procedure. Based on this, the system extracted the path text, relevant video stream slices, and operating condition tags for that operation and added them to the negative sample pool for incremental updates. After the update, when encountering similar emergency modes again, the system output results will be more consistent with the on-site procedures. The purpose of this mechanism is to continuously correct the model and knowledge base using on-site review results, so that the system can maintain its adaptability to the actual on-site operating conditions in long-term operation. In this embodiment, the target area is a gas station, the target object is the operator wearing anti-static work clothes, and the operating procedure text includes the gas leak emergency handling process and valve operation specifications.

[0031] This embodiment provides a practical approach for a specific scenario involving gas stations. Specifically, in this scenario, the target area is limited to the gas station, including areas such as gas unloading, pressure regulation, main valve groups, booster pumps, and pipe corridors. The target personnel are limited to workers wearing anti-static work clothes. The operating procedure text mainly includes gas leak emergency handling procedures and valve operation specifications. Such limitations are not merely for scenario description, but directly affect feature selection and risk interpretation logic in the identification chain. Specifically, since the workers are uniformly dressed, the system does not use clothing color as the only primary feature when extracting appearance features, but pays more attention to information such as body shape, relative body shape, equipment accessories and movement patterns. At the same time, high-risk actions in the station procedures usually revolve around valves, flanges, detectors, isolation zones and emergency response procedures. Therefore, when analyzing video stream slices, the system prioritizes detecting these actions and objects that are directly related to gas safety. For example, whether a portable combustible gas detector is carried out, whether valve opening and closing actions are performed, and whether the evacuation or handling is carried out according to the prescribed route after a leak alarm are all key identification items; the gas leak emergency handling process usually includes a chain constraint of abnormality detection - reporting - wearing protective gear - detection and confirmation - isolation of the area - handling in sequence; valve operation specifications may require confirmation of upstream and downstream status - slow operation - completion of verification - recording time; these rules are significantly different from ordinary plant area inspections, which determines that the system does not generalize to judge whether there is abnormal movement, but focuses on whether the gas process safety sequence is violated when reasoning; In the further fault tolerance mechanism, if the target is not wearing standard anti-static work clothes, such as temporary outsourced personnel or emergency rescue personnel, the system can still include them in the identification scope, but will attach a tag indicating that the personnel attributes do not fully match, to remind subsequent manual confirmation; if the procedure text only provides some clauses, the system can still complete basic anomaly identification, but the compliance judgment will be limited to the scope of the entered clauses; if special equipment is added on site, and the existing valve specifications do not yet cover it, the procedure text and diagrams should be supplemented first, and then the high-level automatic judgment of the relevant area should be enabled. During the joint nighttime maintenance and leak emergency drill at this station, the system continuously tracked multiple workers wearing anti-static work clothes. When worker A completed observation in the main valve group area and entered the booster pump area through the pipe gallery blind spot without sufficient waiting and performed suspected valve operation, the system compared his behavior with the gas leak emergency handling procedure and valve operation specifications item by item, and finally concluded that the gas process sequence was violated and triggered the corresponding level of warning. The purpose of this step is to unify and anchor the identification targets, procedural constraints, and risk interpretations in the methodology to the specific industrial environment of gas stations, thereby achieving engineering applications for real-world safe production scenarios.

[0032] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying abnormal behavior in gas stations by integrating cross-camera behavior chains, characterized in that, include: Step 1: Acquire multiple video streams, historical trajectory data, a 3D spatial topology map, and operation procedure text. The 3D spatial topology map includes blind spot nodes. Step 2: Extract the appearance and motion features of the target object from the multi-channel video streams, construct a directed topology graph based on the three-dimensional spatial topology graph, determine the dynamic tolerance based on the directed topology graph and the historical trajectory data, and generate the cross-camera identity identifier of the target object through multi-dimensional joint matching; The process of generating the cross-camera identity identifier of the target object through multi-dimensional joint matching includes: calculating the topological connectivity probability between nodes in the directed topology graph based on the historical trajectory data; and generating the cross-camera identity identifier through multi-dimensional joint matching using the appearance features, the topological connectivity probability, and the dynamic tolerance. Step 3: Based on the cross-camera identity identifier, stitch together the trajectory of the target object in the field of view corresponding to the multiple video streams and obtain the actual cross-node time. Based on the comparison result of the actual cross-node time and the dynamic tolerance, when the actual cross-node time is greater than the dynamic tolerance, insert the blind spot abnormal stay node corresponding to the blind spot node into the stitched trajectory to generate a spatiotemporal behavior path. Step 4: Based on the spatiotemporal behavior path, extract the multi-channel video streams corresponding to the time period to generate video stream slices. Input the spatiotemporal behavior path, the video stream slices, and the operation procedure text into a visual language model for analysis to generate behavioral intent reasoning results. Convert the spatiotemporal behavior path into a text description, and combine the text description with the video stream slices to construct prompt words. Convert the operation procedure text into a structured knowledge graph, and input the prompt words and the structured knowledge graph into the visual language model for analysis. Step 5: Determine abnormal behavior based on the behavioral intent reasoning results and calculate dynamic risk score, triggering the corresponding graded alarm instruction; The triggering mechanism for the tiered alarm command is as follows: When the dynamic risk score is greater than a preset first risk threshold, a system blocking control signal is generated for the controlled equipment in the gas station and a highest-level alarm data packet is output; when the dynamic risk score is greater than a preset second risk threshold and less than or equal to the preset first risk threshold, a manual review prompt message is generated and a medium-level alarm data packet is output; when the dynamic risk score is less than or equal to the preset second risk threshold, a regular log recording status flag is generated; wherein, the preset first risk threshold is greater than the preset second risk threshold.

2. The method for identifying abnormal behavior in gas stations by fusing cross-camera behavior chains according to claim 1, characterized in that, The multidimensional joint matching performs feature fusion by comparing the similarity of the apparent features, the topological connectivity probability, and the time matching degree determined by comparing the actual cross-node time consumption with the dynamic tolerance, and outputs the best matching object as the cross-camera identity identifier.

3. The method for identifying abnormal behavior in gas stations by fusing cross-camera behavior chains according to claim 1, characterized in that, The abnormal dwell node in the blind zone includes a timestamp and spatial coordinate attributes, which are generated by a preset coordinate mapping in the three-dimensional spatial topology map.

4. The method for identifying abnormal behavior in gas stations by fusing cross-camera behavior chains according to claim 1, characterized in that, The prompt words include an instruction prefix, a contextual behavior description, and visual query content. The contextual behavior description consists of a textual description of the spatiotemporal behavior path, and the visual query content is used to guide the visual language model to recognize the operation actions in the video stream slice.

5. The method for identifying abnormal behavior in gas stations by fusing cross-camera behavior chains according to claim 1, characterized in that, The abnormal behaviors include bypassing to approach critical equipment, unauthorized continuous operation across regions, and abnormal violations in blind spots. The dynamic risk score is obtained by fusing the severity level of the abnormal behavior, the frequency of similar abnormalities within a preset time period, and the duration of abnormal stay nodes in the blind spots.

6. The method for identifying abnormal behavior in gas stations by fusing cross-camera behavior chains according to claim 1, characterized in that, It also includes a knowledge base update step: receiving misjudged labels from manual review feedback, extracting the spatiotemporal behavior path and the video stream slice corresponding to the misjudged label as negative samples, and updating the parameters of the visual language model through a fine-tuning algorithm.

7. The method for identifying abnormal behavior in gas stations by fusing cross-camera behavior chains according to claim 1, characterized in that, The target group is the workers wearing anti-static work clothes, and the operating procedure text includes the emergency handling process for gas leaks and valve operation specifications.

Images