An edge intelligence-based low-light video behavior analysis method

Abstract

The application discloses a low-light video behavior analysis method based on edge intelligence, comprising the following steps: receiving low-light video frames and binding imaging state data; calculating local low-light state and generating an observability graph; detecting and tracking personnel and extracting human body part regions; combining the observability graph to construct a violation behavior evidence graph; identifying evidence gaps and triggering camera micro-exposure supplementary sampling; checking supplementary sampling evidence and updating alarm events. The application makes full use of edge AI video analysis and low-light observability closed-loop supplementary sampling technology, realizes oil and petrochemical operation behavior evidence identification, and has the advantages of low false positives, fast response and high alarm credibility.

Description

Technical Field

[0001] This invention relates to the field of video behavior analysis technology, and in particular to a low-light video behavior analysis method based on edge intelligence. Background Technology

[0002] Existing AI video analytics technologies typically utilize target detection and behavior recognition models deployed on edge devices or servers to perform personnel detection, tracking, and safety behavior identification on video streams captured by cameras. This determines whether workers are engaging in violations such as not wearing safety helmets, work clothes, or gloves, or smoking, and generates alarm events when the identification results meet preset alarm conditions. This type of technology can reduce the workload of manual inspections and is suitable for safety supervision scenarios such as oil and petrochemical plants, construction sites, and industrial plants.

[0003] However, in low-light operating areas of the oil and petrochemical industry, video frames captured by cameras are often affected by dark areas, brightness noise, motion blur, reflection spillover, and edge loss, making it difficult to form stable identification evidence for the head, torso, hands, and mouth / nose areas. Existing technologies typically only perform image enhancement, confidence filtering, or duplicate alarm suppression, without determining whether the human body parts meet the imaging conditions for forming behavioral evidence, nor do they trigger camera micro-exposure supplementary acquisition based on specific evidence gaps when evidence is insufficient, which easily leads to false alarms, missed alarms, and duplicate alarms.

[0004] Therefore, how to provide a low-light video behavior analysis method based on edge intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a low-light video behavior analysis method based on edge intelligence. This invention fully utilizes edge AI video analysis and low-light observability closed-loop supplementary mining technology to achieve evidence-based identification of oil and petrochemical operation behaviors, and has the advantages of low false alarms, fast response and high alarm credibility.

[0006] A low-light video behavior analysis method based on edge intelligence according to an embodiment of the present invention includes the following steps: The edge analysis node deployed in the low-light operation area of ​​the oil and petrochemical industry receives continuous video frames collected by the camera and reads the corresponding imaging status data to obtain the low-light video data to be analyzed. Based on the low-light video data to be analyzed, the brightness noise, edge sharpness, motion blur, reflection spillover, and dark area flooding status of each local area are calculated to generate a low-light observability map. Personnel detection and tracking are performed on low-light video data to be analyzed, and the personnel target tracking results are obtained. Based on the personnel target tracking results, human body parts regions are extracted, including the head region, torso region, hand region, and mouth and nose region. Based on the low-light observability map and human body parts, construct behavioral evidence maps corresponding to the behaviors of not wearing a safety helmet, not wearing anti-static work clothes, not wearing anti-static gloves, and smoking. In the behavioral evidence graph, evidence nodes that cannot form valid evidence are identified, evidence gap information is generated, and edge analysis nodes generate camera micro-exposure re-sampling instructions based on the evidence gap information to obtain re-sampling low-light video frames. The consistency of detail fidelity representation and color fidelity representation of evidence nodes in the supplemented low-light video frames is verified to obtain valid evidence nodes, update the evidence conservation state, and generate or update the corresponding edge AI video analysis alarm events.

[0007] Optionally, the generation of the low-light video data to be analyzed specifically includes: A frame-level synchronization relationship is established between the edge analysis node and the camera, and each consecutive video frame is bound to the imaging status data at the same acquisition time to obtain frame-level bound data. Perform timestamp continuity verification and imaging status integrity verification on the frame-level bound data, and remove consecutive video frames that cannot be bound to the imaging status data to obtain valid frame-level bound data; The original continuous video frames, along with the exposure time, sensor gain, and frame rate bound to the original continuous video frames, are retained from the effective frame-level binding data to obtain the basic data for low light determination. Valid frame-level binding data is obtained by filtering the original continuous video frames in the low-light determination base data where the dark area submersion state reaches a preset low-light ratio threshold or the brightness noise reaches a preset low-light noise threshold, thus obtaining the low-light video data to be analyzed.

[0008] Optionally, the generation of the low-light observability map specifically includes: The continuous video frames in the low-light video data to be analyzed are converted into a luminance channel map. The luminance channel map is then divided into multiple non-overlapping local regions according to a preset pixel size to obtain a sequence of local regions. For each local region in the local region sequence, calculate the brightness noise, edge sharpness, motion blur, reflection overflow and dark area flooding state to obtain the low illumination state of the local region. Low-light observability gating fusion is performed based on the low-light status of local regions to obtain the low-light observability value of each local region. The low-light observability value is backfilled according to the position of each local region in the local region sequence in consecutive video frames, and then bound to the low-light video data to be analyzed to generate a low-light observability map.

[0009] Optionally, the generation of the human body part region specifically includes: Scale normalization and inter-frame order preservation processing are performed on the original continuous video frames in the low-light video data to be analyzed to obtain the video frame sequence to be detected. Input the video frame sequence to be detected and the low-light observability map into the low-light part extraction model to extract the candidate results of people, candidate results of human key points, candidate results of human parts and cross-frame identity features in each video frame to be detected, and obtain the frame-level person detection results. Based on the candidate personnel results, candidate human body key points, candidate human body parts, and cross-frame identity features in the frame-level personnel detection results, cross-frame association is performed to obtain the personnel target tracking results. Based on the trajectory of the same person in the personnel target tracking results, the trajectory attribution binding is performed on the candidate results of the human body parts of the same person in consecutive video frames to obtain the trajectory-bound human body part region. Perform a consistency check on the human body parts regions bound to the trajectory, and remove regions that are inconsistent with the candidate human body, candidate human body key points, and low-light observability map to obtain the human body parts regions, which are then bound together with the human target tracking results to the low-light video data to be analyzed.

[0010] Optionally, the low-light part extraction model includes a low-light observability embedding layer, a temporal feature extraction layer, a personnel detection branch, a human body key point branch, a human body part region regression branch, a cross-frame identity feature branch, and an observability gated output layer.

[0011] Optionally, the generation of the behavioral evidence graph specifically includes: The low-light observability map is mapped to the head, torso, hand, and mouth / nose regions in the same person target tracking result according to the location of the human body parts, thus obtaining the body part observability result; Based on the observability of human body parts and regions, security object evidence identification is performed to obtain security object evidence identification results. Based on the tracking results of the same person target, the observability results of the location, and the identification results of the security object evidence, an evidence node set indexed by the person's trajectory is established to obtain the personnel evidence node set; Based on the behaviors of not wearing a safety helmet, not wearing anti-static work clothes, not wearing anti-static gloves, and smoking, evidence nodes were selected from the personnel evidence node set to obtain four types of behavioral evidence sub-graphs. For each evidence node in the four types of behavioral evidence subgraphs, the corresponding human body part area, evidence node type, and evidence judgment direction are configured to obtain the evidence node attributes. Based on the evidence node attributes and location observability results, evidence node gating is performed on the four types of behavioral evidence subgraphs to obtain the gated behavioral evidence subgraphs, including evidence nodes that are allowed to form evidence and evidence nodes that are gaps in the evidence to be judged. The behavioral evidence sub-graphs after gating are bound to the same personnel trajectory in the personnel target tracking results to generate corresponding behavioral evidence graphs for behaviors such as not wearing a safety helmet, not wearing anti-static work clothes, not wearing anti-static gloves, and smoking.

[0012] Optionally, the generation of the supplementary low-light video frames specifically includes: Read the evidence nodes of the evidence gap to be judged from the behavioral evidence diagram, and determine the evidence nodes that cannot form valid evidence based on the attributes and location observability of the evidence nodes. Evidence gap information is generated based on evidence nodes that cannot form valid evidence, including the trajectory of the same person, the corresponding violation, the corresponding body part area, the attributes of the evidence node, the observability of the part, the dominant cause of the evidence gap, and the accompanying cause of the evidence gap. Based on the evidence gap information and the exposure time, sensor gain, and frame rate in the low-light video data to be analyzed, a camera micro-exposure re-sampling strategy is generated. The camera micro-exposure re-sampling strategy is converted into a camera micro-exposure re-sampling command, and the edge analysis node sends the camera micro-exposure re-sampling command to the camera; Receive continuous video frames collected by the camera according to the camera's micro-exposure re-collection command, bind them to the corresponding evidence node that cannot form valid evidence according to the evidence node write-back identifier, and obtain the re-collected low-light video frames.

[0013] Optionally, the evidence node that cannot form valid evidence is an evidence node that meets the following conditions: the evidence node belongs to the violation support type, violation rebuttal type, or conflict exclusion type, and the observability result of the corresponding human body part area of ​​the evidence node does not reach the preset evidence node observability threshold of the evidence node. When the conflict exclusion type evidence node of smoking behavior has formed the result of walkie-talkie existence, water cup existence, or face wiping action existence, the hand-mouth approach node, the thin and long clamping object existence node, and the smoke or bright spot existence node of the same smoking behavior are not determined as evidence nodes that cannot form valid evidence.

[0014] Optionally, the generation of the edge AI video analysis alarm event specifically includes: Based on the evidence node write-back identifier, the supplementary low-light video frames are located to the same person's trajectory, corresponding violation, and corresponding evidence node in the behavioral evidence map, thus obtaining the supplementary evidence node; Extract edge contours, texture continuity, and small target morphology from the human body parts corresponding to the supplementary evidence nodes to generate detailed and high-fidelity representations. Extract color distribution, material reflection, and color stability after exposure compensation from the human body parts corresponding to the supplementary evidence nodes to generate a color fidelity characterization. Perform a consistency check on the detail fidelity representation and the color fidelity representation to obtain the consistency check result; Based on the consistency verification results, the supplementary evidence nodes are converted into valid evidence nodes, and the valid evidence score is calculated. Update the evidence conservation state based on valid evidence nodes, and generate or update edge AI video analysis alarm events based on the updated evidence conservation state.

[0015] The beneficial effects of this invention are: This invention utilizes edge analysis nodes to locally process continuous video frames and imaging status data in low-light operating areas of the oil and petrochemical industry. This generates a low-light observability map and transforms brightness noise, edge sharpness, motion blur, reflection spillover, and dark area submersion into imaging gating criteria before evidence formation. This allows the determination of behaviors such as not wearing a safety helmet, not wearing anti-static work clothes, not wearing anti-static gloves, and smoking to no longer rely solely on the confidence level of a single frame. This enables the differentiation between situations where "violation characteristics definitely exist" and situations where "they are not clearly visible due to low light," reducing the probability of misjudging unobservable areas as violations.

[0016] This invention further combines low-light observability maps with human body part regions and behavioral evidence maps. Through evidence node gating, it determines evidence nodes that are allowed to form evidence and evidence nodes with gaps in evidence to be judged. When evidence is insufficient, it generates camera micro-exposure re-extraction instructions based on the evidence gap information. This method transforms camera re-extraction from ordinary automatic exposure facing the entire image to targeted re-extraction of the same person's trajectory, corresponding violation behavior, and corresponding human body part region. This improves the identifiability of the head, torso, hands, and mouth / nose regions under low light conditions and reduces invalid calculations and invalid video uploads at the edge.

[0017] This invention also transforms evidence nodes in re-acquired low-light video frames into valid evidence nodes through consistency verification of detail fidelity representation and color fidelity representation. These valid evidence nodes are then used to update the evidence conservation state, generating or updating edge AI video analysis alarm events. This processing can suppress the interference of false edges, false colors, and false textures generated during low-light enhancement or micro-exposure re-acquired images on the recognition results. Simultaneously, by using event fingerprints and evidence conservation states, it reduces repeated alarms for the same person and the same violation, improving the accuracy, real-time performance, and alarm reliability of safety supervision in low-light operation areas of the oil and petrochemical industry. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a low-light video behavior analysis method based on edge intelligence proposed in this invention; Figure 2 This is a flowchart of the evidence gap filling process for a low-light video behavior analysis method based on edge intelligence proposed in this invention. Figure 3 This is a flowchart illustrating the evidence consistency verification process for a low-light video behavior analysis method based on edge intelligence proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figures 1-3 A low-light video behavior analysis method based on edge intelligence includes the following steps: The edge analysis node deployed in the low-light operation area of ​​the oil and petrochemical industry receives continuous video frames collected by the camera and reads the corresponding imaging status data to obtain the low-light video data to be analyzed. Based on the low-light video data to be analyzed, the brightness noise, edge sharpness, motion blur, reflection spillover, and dark area flooding status of each local area are calculated to generate a low-light observability map. Personnel detection and tracking are performed on low-light video data to be analyzed, and the personnel target tracking results are obtained. Based on the personnel target tracking results, human body parts regions are extracted, including the head region, torso region, hand region, and mouth and nose region. Based on the low-light observability map and human body parts, construct behavioral evidence maps corresponding to the behaviors of not wearing a safety helmet, not wearing anti-static work clothes, not wearing anti-static gloves, and smoking. In the behavioral evidence graph, evidence nodes that cannot form valid evidence are identified, evidence gap information is generated, and edge analysis nodes generate camera micro-exposure re-sampling instructions based on the evidence gap information to obtain re-sampling low-light video frames. The consistency of detail fidelity representation and color fidelity representation of evidence nodes in the supplemented low-light video frames is verified to obtain valid evidence nodes, update the evidence conservation state, and generate or update the corresponding edge AI video analysis alarm events.

[0021] In this embodiment, the generation of the low-light video data to be analyzed specifically includes: A frame-level synchronization relationship is established between the edge analysis node and the camera, and each consecutive video frame is bound to the imaging status data at the same acquisition time to obtain frame-level bound data. Perform timestamp continuity verification and imaging status integrity verification on the frame-level bound data, and remove consecutive video frames that cannot be bound to the imaging status data to obtain valid frame-level bound data; The original continuous video frames, along with the exposure time, sensor gain, and frame rate bound to them, are retained from the effective frame-level binding data to obtain the basic data for low-light determination. Valid frame-level binding data is obtained by filtering the original continuous video frames in the low-light determination base data where the dark area submersion state reaches a preset low-light ratio threshold or the brightness noise reaches a preset low-light noise threshold, thus obtaining the low-light video data to be analyzed.

[0022] In this embodiment, the generation of the low-light observability map specifically includes: The continuous video frames in the low-light video data to be analyzed are converted into a luminance channel map. The luminance channel map is then divided into multiple non-overlapping local regions according to a preset pixel size to obtain a sequence of local regions. For each local region in the local region sequence, calculate the brightness noise, edge sharpness, motion blur, reflection overflow and dark area flooding state to obtain the low illumination state of the local region. The brightness noise is calculated by subtracting the brightness value of the brightness channel map and the brightness value of the edge-preserving filter map at each pixel position in the local area. The average value of each difference is calculated by squaring the difference. The square root of the average value is then obtained to get the root mean square of the brightness residual. The root mean square of the brightness residual is divided by a preset low light noise threshold, and the smaller value between the obtained ratio and the first value is taken as the brightness noise of the local area. The edge sharpness calculation involves calculating the brightness gradient of each pixel position in the local area in both the horizontal and vertical directions. The square root of the sum of the squares of the horizontal and vertical brightness gradients is obtained to get the brightness gradient magnitude of that pixel position. The average brightness gradient magnitude of all pixel positions in the local area is then calculated to obtain the average brightness gradient magnitude. The average brightness gradient magnitude is then divided by a preset edge sharpness threshold, and the smaller of the ratio obtained and the threshold value is taken as the edge sharpness of that local area. The motion blur calculation involves calculating the average optical flow amplitude of a local region between adjacent consecutive video frames, dividing the average optical flow amplitude by a preset motion amplitude threshold to obtain the motion amplitude ratio, dividing the preset edge sharpness threshold by the sum of the average brightness gradient amplitude and a preset positive constant to obtain the edge attenuation ratio, and multiplying the motion amplitude ratio and the edge attenuation ratio, taking the smaller value between the product and one as the motion blur of the local region. The reflection overflow statistics include the number of pixels in a local area whose brightness value reaches the preset reflection brightness threshold. The number of pixels is divided by the total number of pixels in the local area to obtain the reflection overflow of that local area. The dark area flooding state is determined by counting the number of pixels in a local area whose brightness value does not reach the preset dark area brightness threshold. This number of pixels is divided by the total number of pixels in the local area to obtain the dark area flooding state of that local area. Low-light observability gating fusion is performed based on the low-light status of local regions to obtain the low-light observability value of each local region. The low-light observability gating fusion uses luminance noise, edge sharpness, motion blur, reflection spillover, and dark area flooding as imaging gating conditions before evidence formation. The low-light observability value of the corresponding local area is determined by the weakest imaging constraint among them. Specifically, luminance noise is subtracted from 1 to obtain the observability component corresponding to luminance noise, edge sharpness is taken as the observable component corresponding to edge sharpness, motion blur is subtracted from 1 to obtain the observable component corresponding to motion blur, reflection spillover is subtracted from 1 to obtain the observable component corresponding to reflection spillover, and dark area flooding is subtracted from 1 to obtain the observable component corresponding to dark area flooding. The minimum value among the five observable components is taken as the low-light observability value of the corresponding local area. The low-light observability value is backfilled according to the position of each local region in the local region sequence in consecutive video frames, and then bound to the low-light video data to be analyzed to generate a low-light observability map.

[0023] In this embodiment, the generation of human body part regions specifically includes: Scale normalization and inter-frame order preservation processing are performed on the original continuous video frames in the low-light video data to be analyzed to obtain the video frame sequence to be detected. Input the video frame sequence to be detected and the low-light observability map into the low-light part extraction model to extract the candidate results of people, candidate results of human key points, candidate results of human parts and cross-frame identity features in each video frame to be detected, and obtain the frame-level person detection results. Based on the candidate personnel results, candidate human body key points, candidate human body parts, and cross-frame identity features in the frame-level personnel detection results, cross-frame association is performed to obtain the personnel target tracking results. Cross-frame association uses cross-frame identity feature similarity, regional overlap of candidate personnel results, displacement of human body center point, spatial consistency of human body key points, and observability consistency of candidate human body parts to jointly determine the association relationship of the same person in adjacent video frames to be detected. Successfully associated candidate personnel results are written into the trajectory of the same person in chronological order to obtain the personnel target tracking result. Based on the trajectory of the same person in the personnel target tracking results, the trajectory attribution binding is performed on the candidate results of the human body parts of the same person in consecutive video frames to obtain the trajectory-bound human body part region. Perform consistency verification on the human body parts regions bound to the trajectory, remove regions that are inconsistent with the candidate human body, candidate human body key points and low light observability map, obtain the human body parts regions, and bind them together with the human target tracking results to the low light video data to be analyzed. The part consistency check restricts the trajectory-bound human part area to the corresponding candidate person result, and determines the spatial assignment of the head area, torso area, hand area, and mouth and nose area according to the candidate human key points. The head area is jointly defined by the head key points and the upper boundary of the candidate person result; the torso area is jointly defined by the shoulder key points, hip key points, and the middle range of the candidate person result; the hand area is defined by the wrist key points and its neighborhood; and the mouth and nose area is defined by the facial key points and its neighborhood. When any trajectory-bound human part area exceeds the corresponding candidate person result, or its spatial assignment is inconsistent with the corresponding candidate human key points, or its low light observability value of the corresponding local area does not reach the observability threshold of the corresponding part, the trajectory-bound human part area is removed from the same person trajectory.

[0024] In this embodiment, the low-light part extraction model includes a low-light observability embedding layer, a temporal feature extraction layer, a personnel detection branch, a human body key point branch, a human body part region regression branch, a cross-frame identity feature branch, and an observability gated output layer. The low-light observability embedding layer downsamples the low-light observability map according to the feature scale of the temporal feature extraction layer, and uses the downsampled low-light observability map as feature gating weights to be superimposed on the shallow texture features and deep semantic features of the video frame sequence to be detected, thus obtaining observability weight features. The temporal feature extraction layer performs intra-frame convolution feature extraction and adjacent frame feature shift fusion on the observability weight features to obtain low-light temporal features. The personnel detection branch outputs personnel candidate results based on the low-light temporal features. The human body keypoint branch outputs human body keypoint candidate results based on the low-light temporal features. The human body part region regression branch regresses the head region, torso region, hand region, and mouth and nose region based on the personnel candidate results and human body keypoint candidate results to obtain human body part candidate results. The cross-frame identity feature branch outputs cross-frame identity features for cross-frame association of the same person based on the low-light temporal features. The observability gating output layer performs validity filtering on the personnel candidate results, human body keypoint candidate results, and human body part candidate results based on the low-light observability map to obtain frame-level personnel detection results. Compared to models that only output person bounding boxes and human body keypoints based on image texture, the low-light part extraction model adds a low-light observability embedding layer and an observability gated output layer. This allows the low-light observability map to participate in the screening of shallow texture features, deep semantic features, and human body part candidate results. The observability gated output layer uses the low-light observability value in the region corresponding to the human body part candidate result as the basis for determining the validity of the part. It suppresses human body part candidate results whose low-light observability value does not reach the corresponding part observability threshold, and retains human body part candidate results whose low-light observability value reaches the corresponding part observability threshold, thus obtaining frame-level person detection results that participate in cross-frame association. The training samples for the low-light location extraction model include video samples from low-light work areas in the oil and petrochemical industry, labeled with personnel bounding boxes, human key points, head region, torso region, hand region, mouth and nose region, and personnel identification tags. The video samples from these low-light work areas are input into the low-light location extraction model, and personnel detection loss, human key point loss, human location region loss, cross-frame identity consistency loss, and observability gating loss are calculated separately. These losses are then weighted and summed to obtain the total model loss. The personnel detection loss is calculated based on the positional deviation and category error between the predicted and labeled personnel bounding boxes; the human key point loss is calculated based on the mean square error between the predicted and labeled key point heatmaps; and the human location region loss is calculated based on the candidate human location data. The model calculates positional deviations, region overlap deviations, and part category errors between the labeled head region, labeled torso region, labeled hand region, and labeled mouth and nose region. The cross-frame identity consistency loss is calculated based on the sorting interval where the cross-frame identity feature distance corresponding to the same person's identity in adjacent video frames is less than the cross-frame identity feature distance corresponding to different person's identity. The observability gating loss is calculated based on the degree to which candidate results for human body parts corresponding to unobservable local regions in the low-light observability map are suppressed, and the degree to which candidate results for human body parts corresponding to observable local regions are preserved. When the decrease in the total model loss on the validation samples does not reach the preset convergence threshold in consecutive preset training rounds, and the recall rate of the person bounding boxes, the accuracy of human body keypoint localization, and the accuracy of human body part region overlap on the validation samples all reach the corresponding preset accuracy thresholds, the low-light part extraction model is considered to have converged.

[0025] In this embodiment, the generation of the behavioral evidence graph specifically includes: The low-light observability map is mapped to the head, torso, hand, and mouth / nose regions in the same person target tracking result according to the location of the human body parts, thus obtaining the body part observability result; The observability results of the body parts are extracted from the low-light observability map according to the position of the body parts in the continuous video frames. The low-light observability values ​​of the corresponding local areas are extracted from the low-light observability map, and the low-light observability values ​​of the same body parts are summarized to obtain the observability results of the head area, the torso area, the hand area, and the mouth and nose area. These results are then linked to the tracking results of the same person target. Based on the observability of human body parts and regions, security object evidence identification is performed to obtain security object evidence identification results. The safety object evidence identification process uses the head, torso, hands, mouth and nose areas, as well as the associated areas between the hands and mouth / nose areas, as identification targets. For the head area, the identification targets include the helmet shell edge, helmet color, and reflective features of the helmet material. For the torso area, the identification targets include the color, texture, reflective strips, and markings of the anti-static work clothes. For the hands area, the identification targets include the edge of the anti-static glove, exposed skin areas, and areas where the material is not covered. For the associated areas between the hands and mouth / nose areas, the identification targets include the proximity of the hands and mouth, long, thin objects being held, smoke, or bright spots. For the hands and adjacent areas, the identification targets include walkie-talkies, water cups, and face-wiping actions. Based on the tracking results of the same person target, the observability results of the location, and the identification results of the security object evidence, an evidence node set indexed by the person's trajectory is established to obtain the personnel evidence node set; The personnel evidence node set includes location observability nodes, target presence nodes, target absence nodes, action relationship nodes, and conflict resolution nodes. Location observability nodes are generated from location observability results. Target presence nodes are generated from the safety object evidence identification results of the presence of the safety helmet shell edge, the anti-static work clothes feature, the anti-static glove cover edge, the slender clamping object, and the smoke or bright spot. Target absence nodes are generated from the safety object evidence identification results of the missing safety helmet shell edge, the missing safety helmet color and material, the missing anti-static work clothes feature, and the missing anti-static glove cover edge. Action relationship nodes are generated from the hand-mouth proximity relationship in the safety object evidence identification results. Conflict resolution nodes are generated from the safety object evidence identification results of the presence of the walkie-talkie, the water cup, and the face-wiping action. Based on the behaviors of not wearing a safety helmet, not wearing anti-static work clothes, not wearing anti-static gloves, and smoking, evidence nodes were selected from the personnel evidence node set to obtain four types of behavioral evidence sub-graphs. The sub-images for evidence of not wearing a safety helmet include observable nodes for the head area, nodes indicating the presence of the head outline, nodes indicating missing edges of the safety helmet shell, and nodes indicating missing color and material of the safety helmet; the sub-images for evidence of not wearing anti-static work clothes include observable nodes for the torso area, nodes indicating the presence of the upper garment area, nodes indicating mismatched color of the anti-static work clothes, nodes indicating mismatched texture of the anti-static work clothes, and nodes indicating missing reflective strips of the anti-static work clothes; the sub-images for evidence of not wearing anti-static gloves include observable nodes for the hand area, nodes indicating the complete outline of the hand, nodes indicating the presence of exposed skin or non-glove material, and nodes indicating missing glove edges; the sub-images for evidence of smoking include observable nodes for the hand area, observable nodes for the mouth and nose area, nodes indicating proximity of the hand and mouth, nodes indicating the presence of a thin, long object, nodes indicating the presence of smoke or bright spots, and nodes indicating conflict resolution. For each evidence node in the four types of behavioral evidence subgraphs, the corresponding human body part area, evidence node type, and evidence judgment direction are configured to obtain the evidence node attributes. The types of evidence nodes include observability gating type, violation support type, violation rebuttal type, and conflict exclusion type. The evidence judgment direction marks the relationship between the evidence nodes and the corresponding behaviors. Among them, the evidence nodes of the violation support type support the generation of valid evidence nodes for the corresponding violation behaviors, the evidence nodes of the violation rebuttal type prevent the increase of the evidence conservation state of the corresponding violation behaviors, and the evidence nodes of the conflict exclusion type prevent the formation of valid evidence nodes for smoking behaviors when a conflict object or conflict action is detected. Based on the evidence node attributes and location observability results, evidence node gating is performed on the four types of behavioral evidence subgraphs to obtain the gated behavioral evidence subgraphs, including evidence nodes that are allowed to form evidence and evidence nodes that are gaps in the evidence to be judged. The evidence node gating system calls the observability results of the corresponding human body part region in the evidence node attributes, and compares the observability results of the corresponding human body part region with the preset observability threshold of the evidence node. When the observability results of the corresponding human body part region reach the preset observability threshold, the evidence node is marked as an evidence node that is allowed to form evidence. When the observability results of the corresponding human body part region do not reach the preset observability threshold, the evidence node is marked as an evidence node with a gap to be judged. The behavioral evidence sub-graphs after gating are bound to the same personnel trajectory in the personnel target tracking results to generate behavioral evidence graphs corresponding to the behaviors of not wearing a safety helmet, not wearing anti-static work clothes, not wearing anti-static gloves, and smoking. The behavioral evidence map uses the trajectory of the same person as an index and retains four types of behavioral evidence sub-maps, evidence node attributes, location observability results, safety object evidence identification results, evidence nodes that can form evidence, and evidence nodes with gaps in evidence to be judged.

[0026] In this embodiment, the generation of supplementary low-light video frames specifically includes: Read the evidence nodes of the evidence gap to be judged from the behavioral evidence diagram, and determine the evidence nodes that cannot form valid evidence based on the attributes and location observability of the evidence nodes. Evidence gap information is generated based on evidence nodes that cannot form valid evidence, including the trajectory of the same person, the corresponding violation, the corresponding body part area, the attributes of the evidence node, the observability of the part, the dominant cause of the evidence gap, and the accompanying cause of the evidence gap. The causes of the primary evidence gap and the accompanying evidence gap are determined based on the contribution of decreased observability in low light within the corresponding human body region. Specifically, this involves extracting the brightness noise, edge sharpness, motion blur, reflection spillover, and dark area submersion status within the human body region corresponding to the evidence node that cannot form valid evidence. Brightness noise is used as the brightness noise contribution, edge sharpness is subtracted to obtain the edge loss contribution, motion blur is used as the motion blur contribution, and reflection spillover is used as the reflection spillover contribution. Dark area submersion status is used as the dark area submersion contribution. The contributions of brightness noise, edge loss, motion blur, reflection spillover, and dark area submersion are then analyzed. The flooding contributions are sorted from largest to smallest. The low light conditions corresponding to the contribution ranked first are determined as the dominant evidence gap cause. The low light conditions corresponding to the contributions ranked second and whose difference from the contribution ranked first does not exceed the preset cause difference range are determined as the accompanying evidence gap cause. When there are multiple evidence nodes in the same person's trajectory that cannot form valid evidence, the priority of supplementary sampling is determined according to the order of smoking behavior, not wearing anti-static gloves, not wearing a safety helmet, and not wearing anti-static work clothes. Evidence nodes with the same corresponding human body part area, the same dominant evidence gap cause, and the same accompanying evidence gap cause are merged into the same evidence gap information. Based on the evidence gap information and the exposure time, sensor gain, and frame rate in the low-light video data to be analyzed, a camera micro-exposure re-sampling strategy is generated. The camera micro-exposure supplementary acquisition strategy uses the corresponding human body part area in the evidence gap information as the supplementary acquisition and metering area. It determines the direction of exposure time adjustment and sensor gain adjustment based on the main evidence gap cause and the accompanying evidence gap cause. During the generation process, the camera micro-exposure supplementary acquisition strategy maintains the frame rate no lower than the frame rate in the low-light video data to be analyzed, and binds the same personnel trajectory, corresponding violation behavior, evidence nodes that cannot form valid evidence, supplementary acquisition and metering area, main adjustment direction and compensation adjustment direction as supplementary acquisition tasks to be executed. When the dominant evidence gap is caused by reflected light overflow, regardless of the cause of the accompanying evidence gap, the exposure time adjustment direction is decreasing, and the sensor gain adjustment direction is maintaining. When the dominant evidence gap is caused by motion blur, regardless of the cause of the accompanying evidence gap, the exposure time adjustment direction is decreasing, and the sensor gain adjustment direction is increasing. When the dominant evidence gap is caused by low illumination conditions corresponding to dark area submersion, brightness noise, or edge loss, and there is no reflected light overflow or motion blur as an accompanying cause of the evidence gap, the exposure time adjustment direction is increasing, and the sensor gain adjustment direction is maintaining. When the dominant evidence gap is caused by low illumination conditions corresponding to dark area submersion, brightness noise, or edge loss, and the accompanying evidence gap is caused by reflected light overflow or motion blur, the exposure time adjustment direction is maintaining, and the sensor gain adjustment direction is increasing. The camera micro-exposure re-sampling strategy is converted into a camera micro-exposure re-sampling command, and the edge analysis node sends the camera micro-exposure re-sampling command to the camera; The camera micro-exposure re-capture instruction includes the re-capture metering area, re-capture frame number, exposure time adjustment direction, sensor gain adjustment direction, frame rate maintenance rules, and evidence node write-back identifier. The evidence node write-back identifier is jointly determined by the same personnel trajectory, corresponding violation behavior, and evidence nodes that cannot form valid evidence. Before sending the camera micro-exposure re-capture instruction, the edge analysis node reads the re-capture execution status of the same camera. When there is an incomplete camera micro-exposure re-capture instruction for the same camera, the new camera micro-exposure re-capture instruction is written into the re-capture queue according to the re-capture priority. When the new camera micro-exposure re-capture instruction and the incomplete camera micro-exposure re-capture instruction have the same evidence node write-back identifier, the incomplete camera micro-exposure re-capture instruction is updated without adding a new re-capture queue item. Receive continuous video frames collected by the camera according to the camera's micro-exposure re-collection command, bind them to the corresponding evidence node that cannot form valid evidence according to the evidence node write-back identifier, and obtain the re-collected low-light video frames.

[0027] In this embodiment, an evidence node that cannot form valid evidence is an evidence node that meets the following conditions: the evidence node belongs to the violation support type, violation rebuttal type, or conflict exclusion type, and the observability result of the corresponding human body part area of ​​the evidence node does not reach the preset evidence node observability threshold of the evidence node. When the conflict exclusion type evidence node of smoking behavior has formed the result of walkie-talkie existence, water cup existence, or face wiping action existence, the hand-mouth approach node, the thin and long clamping object existence node, and the smoke or bright spot existence node of the same smoking behavior are not identified as evidence nodes that cannot form valid evidence.

[0028] In this embodiment, the generation of edge AI video analysis alarm events specifically includes: Based on the evidence node write-back identifier, the supplementary low-light video frames are located to the same person's trajectory, corresponding violation, and corresponding evidence node in the behavioral evidence map, thus obtaining the supplementary evidence node; Extract edge contours, texture continuity, and small target morphology from the human body parts corresponding to the supplementary evidence nodes to generate detailed and high-fidelity representations. The detailed fidelity representation determines the extraction objects according to the evidence node attributes of the supplemented evidence node. When the supplemented evidence node corresponds to not wearing a safety helmet, the upper edge of the head, the arc contour of the brim, and the head contour boundary are extracted from the head area. When the supplemented evidence node corresponds to not wearing anti-static work clothes, the upper garment boundary, reflective strip boundary, seam texture, and marking contour are extracted from the torso area. When the supplemented evidence node corresponds to not wearing anti-static gloves, the covering boundary from the wrist to the fingers, the outer contour of the fingers, and the boundary of the exposed skin area are extracted from the hand area. When the supplemented evidence node corresponds to smoking behavior, the hand-mouth approach trajectory, the contour of the slender clamped object, the smoke diffusion edge, and the bright spot change trajectory are extracted from the hand area, the mouth and nose area, and the associated area between the hand area and the mouth and nose area. The edge contour, texture continuity, and small target shape are encapsulated together as detailed fidelity representation. Extract color distribution, material reflection, and color stability after exposure compensation from the human body parts corresponding to the supplementary evidence nodes to generate a color fidelity characterization. The color fidelity characterization involves performing an exposure compensation inverse transformation on the human body part region corresponding to the supplemented evidence node according to the exposure time and sensor gain retained in the supplemented low-light video frame, and comparing the color distribution after the exposure compensation inverse transformation with the color stability of the human body part region corresponding to the same evidence node in the low-light video data to be analyzed. Specifically, when the supplementary evidence node corresponds to not wearing a safety helmet, the color template matching result of the safety helmet and the continuity of the safety helmet material reflection are extracted; when the supplementary evidence node corresponds to not wearing anti-static work clothes, the color template matching result of the anti-static work clothes, the brightness continuity of the anti-static work clothes reflective strip, and the color stability of the anti-static work clothes marking are extracted; when the supplementary evidence node corresponds to not wearing anti-static gloves, the material color of the anti-static gloves, the color of the exposed skin, and the color of the non-glove material are extracted; when the supplementary evidence node corresponds to smoking behavior, the color change of the bright spot at the end of the slender clamp and the color change of the gray-white diffusion of the smoke are extracted. The color distribution, material reflection, and color stability after exposure compensation are encapsulated together as a color fidelity characterization. Perform a consistency check on the detail fidelity representation and the color fidelity representation to obtain the consistency check result; The consistency verification is based on the evidence node attributes of the supplementary evidence nodes. The detailed representation judgment result is generated by the detailed fidelity representation, and the color representation judgment result is generated by the color fidelity representation. Both the detailed representation judgment result and the color representation judgment result include supporting the violation, refuting the violation, preventing the formation of evidence of smoking behavior, and not forming evidence. Specifically, when both the detail representation judgment result and the color representation judgment result are consistent with the evidence judgment direction of the supplementary evidence node, and the overlap of the target areas corresponding to the two reaches the preset overlap threshold, and the human body parts corresponding to the two reach the corresponding observability threshold, the consistency verification result is marked as passed. When the detail representation judgment result and the color representation judgment result are inconsistent, or any representation judgment result is inconsistent with the evidence judgment direction of the supplementary evidence node, or the overlap of the target areas corresponding to the two does not reach the preset overlap threshold, or the human body parts corresponding to any representation do not reach the corresponding observability threshold, the consistency verification result is marked as failed. For smoking behavior, when the walkie-talkie existence result, water cup existence result, or face wiping action existence result corresponding to the conflict resolution node are valid, the consistency verification results of the hand-mouth approach node, the thin and long clamping object existence node, and the smoke or bright spot existence node are marked as failed. Based on the consistency verification results, the supplementary evidence nodes are converted into valid evidence nodes, and the valid evidence score is calculated. When the consistency check result is passed, the supplementary evidence node is converted into a valid evidence node. When the supplementary evidence node belongs to the violation support type, the valid evidence node is marked as a valid evidence node supporting the violation. When the supplementary evidence node belongs to the violation rebuttal type, the valid evidence node is marked as a valid evidence node rebutting the violation. When the supplementary evidence node belongs to the conflict exclusion type, the valid evidence node is marked as a conflict exclusion valid evidence node. When the consistency check result is failed, the supplementary evidence node is not converted into a valid evidence node, and the corresponding evidence node is kept in the pending evidence gap state in the behavioral evidence graph. The effective evidence score is obtained by averaging the low-light observability values ​​of the human body parts corresponding to the supplementary evidence nodes, obtaining the observability value of the part evidence, dividing the overlapping area of ​​the target area corresponding to the detail fidelity representation and the target area corresponding to the color fidelity representation by the union area of ​​the two, obtaining the representation overlap value, reading the preset evidence node type weight according to the evidence node type to which the supplementary evidence node belongs, and multiplying the part evidence observability value, the representation overlap value and the preset evidence node type weight to obtain the effective evidence score. Update the evidence conservation state based on valid evidence nodes, and generate or update edge AI video analysis alarm events based on the updated evidence conservation state; The evidence conservation state is maintained using the same person's trajectory and corresponding violation as an index. For the same person's trajectory and corresponding violation, the forgetting coefficient decay is first applied to the evidence conservation state at the previous moment. Then, the valid evidence scores of the nodes supporting the violation are added to the evidence conservation state. The valid evidence scores of the nodes refuting the violation and the nodes excluding the conflict are deducted from the evidence conservation state. When the updated evidence conservation state reaches the preset alarm trigger threshold for the corresponding violation, an edge AI video analysis alarm event is generated or updated. When the updated evidence conservation state is lower than the preset alarm cancellation threshold, and no node supporting the violation appears in the consecutive preset number of valid evidence nodes for the same person's trajectory, the corresponding edge AI video analysis alarm event is turned off. The edge AI video analysis alarm event is formed by camera identifier, work area identifier, the same personnel trajectory and corresponding violation behavior to form an event fingerprint. When there is no edge AI video analysis alarm event in the event fingerprint, a new edge AI video analysis alarm event is generated. When there is an edge AI video analysis alarm event in the event fingerprint, the valid evidence nodes, low-light video frames, evidence conservation status and event duration in the edge AI video analysis alarm event are updated.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a safety supervision scenario in the nighttime pump area, loading and unloading platform, and tank area inspection passage of an oil and petrochemical enterprise. In this scenario, workers are required to wear safety helmets, anti-static work clothes, and anti-static gloves during nighttime inspections, valve operations, and material loading and unloading, and smoking is strictly prohibited. The on-site cameras are affected by factors such as shadows from the pipe corridor, insufficient local lighting, glare from metal equipment, interference from vehicle lights, and rapid personnel movement. Head, torso, hands, and mouth / nose areas often appear unclear, have broken edges, color distortion, and ghosting. Traditional AI video analysis easily misjudges an unclear safety helmet as not being worn, a missing edge of anti-static gloves as bare hands, or ordinary actions like bringing hands close to the face as smoking.

[0030] In this scenario, edge analysis nodes are deployed on the camera access side to receive continuous video frames captured by the cameras and simultaneously read imaging status data such as exposure time, sensor gain, and frame rate to generate low-light video data for analysis. The edge analysis nodes first calculate brightness noise, edge sharpness, motion blur, reflection overflow, and dark area submersion in local regions of the video frames to generate a low-light observability map. Then, they extract personnel target tracking results, including head, torso, hand, and mouth / nose regions, using a low-light area extraction model. During the training phase, the low-light part extraction model uses on-site nighttime video samples and low-light perturbation samples to label personnel bounding boxes, human key points, human part regions, personnel identification marks, and evidence objects such as safety helmets, anti-static work clothes, anti-static gloves, hand-held objects, smoke, or bright spots. During training, the low-light observability map is embedded with shallow texture features and deep semantic features. The model is constrained by personnel detection loss, human key point loss, human part region loss, cross-frame identity consistency loss, and observability gating loss, so that the model prioritizes the preservation of human part regions that can form evidence in low-light and reflective environments.

[0031] During on-site operation, edge analysis nodes construct behavioral evidence maps based on low-light observability maps and human body region data. This transforms identification results such as missing safety helmet edges, missing reflective strips on anti-static work clothes, missing glove edges, hand-mouth contact, long, thin objects being held, smoke or bright spots, and actions like walkie-talkies, water cups, and face-wiping into evidence nodes. When the observability of the human body region corresponding to a certain evidence node is insufficient, a violation conclusion is not directly formed. Instead, evidence gap information is generated, and a micro-exposure re-capture command is sent to the camera based on the cause of the evidence gap. After the re-captured low-light video frames are written back to the corresponding evidence node, the edge analysis node extracts detail fidelity representation and color fidelity representation. Only when the two judgments are consistent and the overlap relationship of the corresponding target areas meets the evidence formation conditions is the re-captured evidence node converted into a valid evidence node, and the evidence conservation state is updated. Ultimately, an edge AI video analysis alarm event is generated or updated.

[0032] On-site operation records show that during nighttime inspections, loading and unloading operations, and tank area patrols, this invention can distinguish between images where "evidence cannot be formed under low light" and images where "evidence of violations is already present." In cases where the head is obscured by shadows, hands are affected by shadows, work clothes are interfered with by reflections, or hand and mouth movements are interfered with by walkie-talkies or water cups, the edge analysis node does not immediately output a violation alarm. Instead, it first triggers micro-exposure re-sampling and dual-characterization consistency verification. Only when the detailed structure and color information after re-sampling support the same violation conclusion is a valid evidence node formed and the alarm event updated. Therefore, the alarms received by on-site safety personnel include personnel trajectory, work area, re-sampling low-light video frames, valid evidence nodes, and evidence conservation status, facilitating verification of the alarm source and reducing the impact of false alarms, missed alarms, and duplicate alarms under low light conditions on nighttime safety supervision in the oil and petrochemical industry.

[0033] Table 1. Performance Comparison of Behavioral Analysis in Low-Light Operation Areas of Petroleum and Petrochemical Industries

[0034] The above comparison was conducted in a video behavior analysis scenario in a low-light working area of ​​the oil and petrochemical industry. The identified objects included not wearing a safety helmet, not wearing anti-static work clothes, not wearing anti-static gloves, and smoking. Table 1 shows that the YOLOv8n+ByteTrack+fixed confidence alarm method has a short response time at the edge, but it directly relies on the appearance features and detection confidence in the video frame. When there is significant dark area coverage, reflection overflow, and motion blur, it easily misidentifies "unclear" head, hand, or mouth / nose areas as violations, resulting in a high average number of false alarms and a high proportion of duplicate alarms. After Zero-DCE enhancement, the recognition accuracy of safety helmets, anti-static work clothes, anti-static gloves, and smoking behavior all improved. However, image enhancement also amplifies local noise and pseudo-textures, and there are still many false alarms under metallic reflections and vehicle headlight interference.

[0035] The SNR-Aware+YOLOv8n+ByteTrack+multi-frame voting alarm method further reduces the number of false alarms and the proportion of duplicate alarms compared to the previous two methods, indicating that performing low-light enhancement based on the signal-to-noise ratio and adopting a multi-frame confirmation strategy can improve the recognition stability in low-light scenes. However, this method still uses the enhanced video frames as the main judgment basis, without distinguishing whether the human body parts have the imaging conditions to form behavioral evidence, and without supplementing the imaging of the corresponding human body parts when the evidence is insufficient. Therefore, in tasks with small targets, weak textures, and strong temporal correlation, such as antistatic gloves and smoking behavior, the improvement in recognition accuracy is limited, and the average number of missed alarms remains at a high level.

[0036] The method of this invention achieves higher recognition accuracy across four types of behaviors compared to the comparative methods, with significantly reduced average false alarms, average missed alarms, and the proportion of duplicate alarms. This is because the invention first determines whether a local area meets the conditions for evidence formation using a low-light observability map. Then, it organizes the human body part area, the evidence recognition results of the safe object, and the observability results of the part into a behavioral evidence map, preventing unclear areas from directly leading to a conclusion of violation. When an evidence node cannot form valid evidence, the edge analysis node generates a camera micro-exposure re-sampling instruction based on the evidence gap information, performing targeted re-sampling of the head, torso, hand, or mouth / nose areas, thereby improving the identifiability of key behavioral evidence under low light conditions. Before the re-sampling of low-light video frames enters subsequent judgment, it also needs to undergo consistency verification of detail fidelity representation and color fidelity representation, which can suppress interference from false edges, false colors, and false textures caused by low-light enhancement or exposure adjustment. Although the average alarm response time of the method of this invention is increased compared with the basic detection method, the increased time mainly comes from evidence gap judgment, micro-exposure re-sampling and consistency verification. In return, it results in lower false alarms, missed alarms and duplicate alarms, and has higher alarm credibility and on-site verification value in the safety supervision scenario of low light operation area of ​​oil and petrochemical industry.

[0037] As can be seen from Table 1, this invention does not simply rely on low-light enhancement to improve image brightness. Instead, it forms a closed-loop analysis process through low-light observability maps, behavioral evidence maps, evidence gap information, camera micro-exposure re-sampling instructions, and dual-representation consistency verification. This transforms edge AI video analysis from single-frame confidence judgment to evidence chain judgment, thereby improving the stability of violation behavior identification, reducing invalid and duplicate alarms, and enhancing the interpretability and credibility of alarm events in the oil and petrochemical operation environment where low light, reflection, ghosting, and partial occlusion coexist.

[0038] The above are merely preferred embodiments 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 low-light video behavior analysis method based on edge intelligence, characterized in that, Includes the following steps: The edge analysis node deployed in the low-light operation area of ​​the oil and petrochemical industry receives continuous video frames collected by the camera and reads the corresponding imaging status data to obtain the low-light video data to be analyzed. Based on the low-light video data to be analyzed, the brightness noise, edge sharpness, motion blur, reflection spillover, and dark area flooding status of each local area are calculated to generate a low-light observability map. Personnel detection and tracking are performed on low-light video data to be analyzed, and the personnel target tracking results are obtained. Based on the personnel target tracking results, human body parts regions are extracted, including the head region, torso region, hand region, and mouth and nose region. Based on the low-light observability map and human body parts, construct behavioral evidence maps corresponding to the behaviors of not wearing a safety helmet, not wearing anti-static work clothes, not wearing anti-static gloves, and smoking. In the behavioral evidence graph, evidence nodes that cannot form valid evidence are identified, evidence gap information is generated, and edge analysis nodes generate camera micro-exposure re-sampling instructions based on the evidence gap information to obtain re-sampling low-light video frames. The consistency of detail fidelity representation and color fidelity representation of evidence nodes in the supplemented low-light video frames is verified to obtain valid evidence nodes, update the evidence conservation state, and generate or update the corresponding edge AI video analysis alarm events.

2. The low-light video behavior analysis method based on edge intelligence according to claim 1, characterized in that, The generation of the low-light video data to be analyzed specifically includes: A frame-level synchronization relationship is established between the edge analysis node and the camera, and each consecutive video frame is bound to the imaging status data at the same acquisition time to obtain frame-level bound data. Perform timestamp continuity verification and imaging status integrity verification on the frame-level bound data, and remove consecutive video frames that cannot be bound to the imaging status data to obtain valid frame-level bound data; The original continuous video frames, along with the exposure time, sensor gain, and frame rate bound to the original continuous video frames, are retained from the effective frame-level binding data to obtain the basic data for low light determination. Valid frame-level binding data is obtained by filtering the original continuous video frames in the low-light determination base data where the dark area submersion state reaches a preset low-light ratio threshold or the brightness noise reaches a preset low-light noise threshold, thus obtaining the low-light video data to be analyzed.

3. The low-light video behavior analysis method based on edge intelligence according to claim 1, characterized in that, The generation of the low-light observability map specifically includes: The continuous video frames in the low-light video data to be analyzed are converted into a luminance channel map. The luminance channel map is then divided into multiple non-overlapping local regions according to a preset pixel size to obtain a sequence of local regions. For each local region in the local region sequence, calculate the brightness noise, edge sharpness, motion blur, reflection overflow and dark area flooding state to obtain the low illumination state of the local region. Low-light observability gating fusion is performed based on the low-light status of local regions to obtain the low-light observability value of each local region. The low-light observability value is backfilled according to the position of each local region in the local region sequence in consecutive video frames, and then bound to the low-light video data to be analyzed to generate a low-light observability map.

4. The low-light video behavior analysis method based on edge intelligence according to claim 1, characterized in that, The generation of the aforementioned human body region specifically includes: Scale normalization and inter-frame order preservation processing are performed on the original continuous video frames in the low-light video data to be analyzed to obtain the video frame sequence to be detected. Input the video frame sequence to be detected and the low-light observability map into the low-light part extraction model to extract the candidate results of people, candidate results of human key points, candidate results of human parts and cross-frame identity features in each video frame to be detected, and obtain the frame-level person detection results. Based on the candidate personnel results, candidate human body key points, candidate human body parts, and cross-frame identity features in the frame-level personnel detection results, cross-frame association is performed to obtain the personnel target tracking results. Based on the trajectory of the same person in the personnel target tracking results, the trajectory attribution binding is performed on the candidate results of the human body parts of the same person in consecutive video frames to obtain the trajectory-bound human body part region. Perform a consistency check on the human body parts regions bound to the trajectory, and remove regions that are inconsistent with the candidate human body, candidate human body key points, and low-light observability map to obtain the human body parts regions, which are then bound together with the human target tracking results to the low-light video data to be analyzed.

5. The low-light video behavior analysis method based on edge intelligence according to claim 1, characterized in that, The low-light part extraction model includes a low-light observability embedding layer, a temporal feature extraction layer, a personnel detection branch, a human body key point branch, a human body part region regression branch, a cross-frame identity feature branch, and an observability gated output layer.

6. The low-light video behavior analysis method based on edge intelligence according to claim 1, characterized in that, The generation of the behavioral evidence graph specifically includes: The low-light observability map is mapped to the head, torso, hand, and mouth / nose regions in the same person target tracking result according to the location of the human body parts, thus obtaining the body part observability result; Based on the observability of human body parts and regions, security object evidence identification is performed to obtain security object evidence identification results. Based on the tracking results of the same person target, the observability results of the location, and the identification results of the security object evidence, an evidence node set indexed by the person's trajectory is established to obtain the personnel evidence node set; Based on the behaviors of not wearing a safety helmet, not wearing anti-static work clothes, not wearing anti-static gloves, and smoking, evidence nodes were selected from the personnel evidence node set to obtain four types of behavioral evidence sub-graphs. For each evidence node in the four types of behavioral evidence subgraphs, the corresponding human body part area, evidence node type, and evidence judgment direction are configured to obtain the evidence node attributes. Based on the evidence node attributes and location observability results, evidence node gating is performed on the four types of behavioral evidence subgraphs to obtain the gated behavioral evidence subgraphs, including evidence nodes that are allowed to form evidence and evidence nodes that are gaps in the evidence to be judged. The behavioral evidence sub-graphs after gating are bound to the same personnel trajectory in the personnel target tracking results to generate corresponding behavioral evidence graphs for behaviors such as not wearing a safety helmet, not wearing anti-static work clothes, not wearing anti-static gloves, and smoking.

7. The low-light video behavior analysis method based on edge intelligence according to claim 1, characterized in that, The generation of the supplementary low-light video frames specifically includes: Read the evidence nodes of the evidence gap to be judged from the behavioral evidence diagram, and determine the evidence nodes that cannot form valid evidence based on the attributes and location observability of the evidence nodes. Evidence gap information is generated based on evidence nodes that cannot form valid evidence, including the trajectory of the same person, the corresponding violation, the corresponding body part area, the attributes of the evidence node, the observability of the part, the dominant cause of the evidence gap, and the accompanying cause of the evidence gap. Based on the evidence gap information and the exposure time, sensor gain, and frame rate in the low-light video data to be analyzed, a camera micro-exposure re-sampling strategy is generated. The camera micro-exposure re-sampling strategy is converted into a camera micro-exposure re-sampling command, and the edge analysis node sends the camera micro-exposure re-sampling command to the camera; Receive continuous video frames collected by the camera according to the camera's micro-exposure re-collection command, bind them to the corresponding evidence node that cannot form valid evidence according to the evidence node write-back identifier, and obtain the re-collected low-light video frames.

8. The low-light video behavior analysis method based on edge intelligence according to claim 7, characterized in that, The evidence nodes that cannot form valid evidence are evidence nodes that meet the following conditions: the evidence node belongs to the violation support type, violation rebuttal type, or conflict exclusion type, and the observability result of the corresponding human body part area of ​​the evidence node does not reach the preset evidence node observability threshold of the evidence node. When the conflict exclusion type evidence node of smoking behavior has formed the result of the existence of a walkie-talkie, the existence of a water cup, or the existence of the face wiping action, the hand-mouth approach node, the thin and long clamping object existence node, and the smoke or bright spot existence node of the same smoking behavior are not identified as evidence nodes that cannot form valid evidence.

9. The low-light video behavior analysis method based on edge intelligence according to claim 1, characterized in that, The generation of the edge AI video analysis alarm event specifically includes: Based on the evidence node write-back identifier, the supplementary low-light video frames are located to the same person's trajectory, corresponding violation, and corresponding evidence node in the behavioral evidence map, thus obtaining the supplementary evidence node; Extract edge contours, texture continuity, and small target morphology from the human body parts corresponding to the supplementary evidence nodes to generate detailed and high-fidelity representations. Extract color distribution, material reflection, and color stability after exposure compensation from the human body parts corresponding to the supplementary evidence nodes to generate a color fidelity characterization. Perform a consistency check on the detail fidelity representation and the color fidelity representation to obtain the consistency check result; Based on the consistency verification results, the supplementary evidence nodes are converted into valid evidence nodes, and the valid evidence score is calculated. Update the evidence conservation state based on valid evidence nodes, and generate or update edge AI video analysis alarm events based on the updated evidence conservation state.

Images