Security inspection method and system based on edge intelligence and dynamic risk assessment, and medium

Abstract

The application discloses a kind of based on edge intelligence and dynamic risk assessment safety inspection method, system and medium, it is related to dynamic risk technical field, including: generating high-altitude construction safety model;Safety point array of high-altitude worker is constructed;Generate high-altitude worker initial information;Enhanced time series point array is generated;Obtain high-altitude worker motion trajectory;Generate high-altitude worker state information;Generate high-altitude construction dynamic risk result, determine safety risk grade and generate high-altitude construction safety inspection response instruction, realize high-altitude construction area safety inspection.The application improves the time series continuity of risk identification and determination accuracy, enhances the reliability of high-altitude construction safety inspection.

Description

Technical Field

[0001] This invention relates to the field of dynamic risk technology, and in particular to a security inspection method, system and medium based on edge intelligence and dynamic risk assessment. Background Technology

[0002] With the continuous expansion of high-altitude construction scale and the increasing complexity of working environments, high-altitude operation safety inspection technology based on video surveillance has gradually become an important means of ensuring construction safety. In high-altitude construction scenarios, due to the large range of personnel movements, the complex structure of safety ropes, and the fact that anchor points are mostly fixed but easily obstructed structures, traditional methods have weak stable tracking capabilities in continuous video frames, making it difficult to obtain continuous and reliable key point motion information.

[0003] Existing methods based on general target tracking or key point tracking models typically rely solely on visual features for cross-frame association, lacking the ability to model construction safety zone constraints and safety rope connection relationships. This leads to key point tracking easily drifting or being lost under conditions of occlusion, lighting changes, or background interference. Regarding the connection status between the safety rope hook point and anchor point, existing technologies mostly employ static distance judgment or single-frame rule analysis, lacking a dynamic analysis mechanism based on continuous trajectories, making it difficult to accurately reflect the connection stability of the safety rope during actual operations.

[0004] In terms of risk assessment, existing technologies typically treat abnormal human posture and the risk of safety rope detachment separately, lacking a unified dynamic risk quantification model. This makes it difficult to comprehensively consider the impact of multiple risk factors on the overall safety status, resulting in insufficient stability and reliability of risk assessment results. Summary of the Invention

[0005] One objective of this invention is to propose a safety inspection method, system, and medium based on edge intelligence and dynamic risk assessment. This invention improves the temporal continuity and accuracy of risk identification, and enhances the reliability of safety inspections for high-altitude construction.

[0006] A security inspection method based on edge intelligence and dynamic risk assessment according to an embodiment of the present invention includes:

[0007] Collect video inspection image sequences of the target high-altitude construction area to generate a high-altitude construction safety model;

[0008] Based on the high-altitude construction safety model, the high-altitude workers in the video inspection image sequence are located, and key sensing parts of the high-altitude workers' body structure and safety protection structure are identified to construct a high-altitude worker safety matrix.

[0009] Based on the safety dot matrix of high-altitude workers, dense feature points are initialized for continuous video frames in the video inspection image sequence. The human body joint feature points, safety rope buckle points and safety anchor position feature points are converted into dense dot matrix feature representations. The dense dot matrix feature representations are then input into the improved CoTracker tracking network to generate initial information of high-altitude workers.

[0010] The initial information of the high-altitude workers is input into the CoTracker temporal sliding module, which performs cross-frame correlation calculation on the dense dot matrix feature representation in continuous video frames and dynamically adjusts the dot feature correlation weights to generate an enhanced temporal dot matrix.

[0011] Based on the enhanced temporal matrix, the improved CoTracker motion update module is used to continuously predict the trajectory of personnel joint feature points, safety rope hook points and safety anchor position feature points to obtain the movement trajectory of high-altitude workers.

[0012] Based on the movement trajectory of the high-altitude workers, the dynamic analysis of the posture changes and safety rope connection status of the high-altitude workers in continuous video frames is performed to generate the status information of the high-altitude workers.

[0013] Based on the status information of high-altitude workers, a dynamic risk assessment model for high-altitude construction is constructed to quantify the risks of abnormal posture and safety rope detachment, generate dynamic risk results for high-altitude construction, determine the safety risk level, and generate high-altitude construction safety inspection response instructions to realize safety inspection of high-altitude construction areas.

[0014] Optionally, generating the high-altitude construction safety model includes:

[0015] Collect continuous video frames generated by fixed monitoring equipment or inspection monitoring equipment in the target high-altitude construction area, and generate a video inspection image sequence.

[0016] Based on the construction structure information of the high-altitude construction area, the video inspection image sequence is spatially calibrated to generate the high-altitude construction area calibration.

[0017] Based on the marking of the high-altitude construction area, the information on the activity range of the workers, and the safety protection requirements, determine the safe operation boundary area, the safety rope connection and restraint area, and the high-altitude fall risk area under the spatial location of the high-altitude construction area.

[0018] The safe operation boundary area, the safety rope connection constraint area, and the high-altitude fall risk area are projected onto the video inspection image frames in the video inspection image sequence to generate a safe operation pixel mask, a safety rope connection pixel mask, and a high-altitude fall pixel mask.

[0019] The safety operation boundary area, safety rope connection constraint area, high-altitude fall risk area, safety operation pixel mask, safety rope connection pixel mask and high-altitude fall pixel mask are associated and encapsulated to generate a high-altitude construction safety model.

[0020] Optionally, the step of locating high-altitude workers in the video inspection image sequence based on the high-altitude construction safety model includes:

[0021] Based on the safety operation pixel mask, safety rope connection pixel mask and high-altitude fall pixel mask in the high-altitude construction safety model, spatial constraint screening is performed on the video inspection image frames to determine the target positioning area of ​​high-altitude workers.

[0022] Within the target positioning area of ​​the high-altitude workers, the high-altitude workers are located, and the high-altitude workers positioning results are generated.

[0023] Based on the positioning results of high-altitude workers, key sensory parts of the high-altitude workers' body structure are identified to obtain the human joint feature points of the workers.

[0024] Based on the positioning results of high-altitude workers and the pixel mask of safety rope connection in the high-altitude construction safety model, the key sensing parts of the safety protection structure of high-altitude workers are identified to obtain the safety rope hook points.

[0025] Based on the safety rope connection pixel mask and the high-altitude construction area calibration in the high-altitude construction safety model, the key sensing parts of the anchoring structure corresponding to the fixed position of the safety rope are identified to obtain the safety anchoring position feature points.

[0026] Based on the spatial relationship between the joint feature points of the human body, the safety rope buckle points, and the safety anchor position feature points, the key sensing parts corresponding to the same high-altitude worker are associated and grouped to construct a safety point matrix for high-altitude workers.

[0027] Optionally, the step of initializing dense feature points in consecutive video frames of the video inspection image sequence based on the safety dot matrix for high-altitude workers includes:

[0028] Based on the safety point matrix for high-altitude workers, determine the human joint feature points, safety rope hook points, and safety anchor position feature points corresponding to the same high-altitude worker, as well as the initial range of the corresponding dense feature points.

[0029] Neighborhood sampling is performed within the initial range of dense feature points of human joint feature points, safety rope buckle points, and safety anchor position feature points to generate dense sets of human joint points, safety rope buckle points, and safety anchor points.

[0030] The dense set of human joints, the dense set of safety rope buckles, and the dense set of safety anchors are merged to generate a dense set of high-altitude workers, and a dense lattice feature representation with spatial topological relationships is generated.

[0031] The dense lattice feature representation with spatial topological relationships is input into the improved CoTracker tracking network to generate spatial visual features of continuous video frames.

[0032] Based on the spatial visual features of continuous video frames, feature sampling is performed on each dense point in the set of dense points of high-altitude workers to generate initial information of high-altitude workers.

[0033] Optionally, inputting the initial information of the high-altitude workers into the CoTracker time-series sliding module includes:

[0034] Based on the initial information of the high-altitude workers, and according to the video frame acquisition time of the video inspection image sequence, a time-series sliding window is constructed to input information;

[0035] Based on the input information from the temporal sliding window, the CoTracker temporal sliding module performs visual similarity calculations on dense points in consecutive video frames to generate initial point feature association weights.

[0036] Based on the safe operation boundary area, the safety rope connection constraint area, and the high-altitude fall risk area, determine the physical safety constraint coefficient corresponding to each dense point in the time-series sliding window input information;

[0037] Based on the spatial topological relationships and point type identifiers in the initial information of high-altitude workers, determine the topological constraint coefficients between the dense points to be associated and the candidate dense points to be associated;

[0038] The initial point feature association weights are dynamically adjusted based on the physical security constraint coefficients and topological constraint coefficients to generate dynamically adjusted point feature association weights.

[0039] Based on the dynamically adjusted point feature association weights, the spatial visual feature vectors corresponding to candidate associated dense points in the temporal sliding window input information are weighted and fused to generate enhanced dense point temporal feature vectors and obtain an enhanced temporal point matrix.

[0040] Optionally, the step of predicting the continuous trajectory of human joint feature points, safety rope hook points, and safety anchor position feature points using the improved CoTracker motion update module based on the enhanced temporal point matrix includes:

[0041] Based on the enhanced timing point matrix, determine the enhanced timing points of human joints, safety rope buckles, and safety anchors for the same high-altitude worker;

[0042] Input the enhanced timing points of human joints, safety rope buckles, and safety anchors into the improved CoTracker motion update module to obtain a set of displacement increments of the enhanced timing points.

[0043] Based on the enhanced temporal point displacement increment set, the pixel position of each enhanced temporal point in the video inspection image sequence is continuously updated to generate an enhanced prediction trajectory set.

[0044] Based on the enhanced predicted trajectory set and spatial topological relationships, the enhanced continuous predicted trajectories originating from the same person's human joint feature points are aggregated to generate human joint feature point trajectories.

[0045] Based on the enhanced predicted trajectory set and spatial topological relationships, the enhanced continuous predicted trajectories originating from the same safety rope hook point are aggregated to generate the motion trajectory of the safety rope hook point.

[0046] Based on the enhanced prediction trajectory set and spatial topological relationships, the enhanced continuous prediction trajectories originating from the same safe anchorage location feature point are spatially stabilized to generate the safe anchorage location feature point trajectory.

[0047] The trajectory of human joint feature points, the movement trajectory of safety rope buckle points, and the trajectory of safety anchor position feature points are associated with the same high-altitude worker to generate the movement trajectory of the high-altitude worker.

[0048] Optionally, the step of dynamically analyzing the posture changes and safety rope connection status of the high-altitude worker in continuous video frames based on the worker's movement trajectory includes:

[0049] Based on the trajectory of human joint feature points, the movement trajectory of safety rope hook points, and the trajectory of safety anchor position feature points, the trajectory is aligned according to the video frame acquisition time to generate a trajectory sequence for analyzing the status of high-altitude workers.

[0050] Based on the trajectory of human joint feature points, calculate the geometric features of human posture of high-altitude workers in continuous video frames, and generate a sequence of posture geometric features.

[0051] Based on the sequence of posture geometric features, the state of human posture changes of high-altitude workers in continuous video frames is determined, and the human posture change state is generated.

[0052] Based on the motion trajectory of the safety rope hook point and the trajectory of the safety anchor position feature point, calculate the relative motion characteristics between the safety rope hook point and the safety anchor position feature point, and generate a safety rope relative connection feature sequence.

[0053] Based on the relative connection feature sequence of the safety rope, determine the connection relationship between the safety rope hook point and the safety anchor position feature point corresponding to the same high-altitude worker, and generate the safety rope connection status.

[0054] Based on the changes in human posture and the connection status of the safety rope, and combined with the trajectory sequence of the high-altitude worker's status analysis, the high-altitude worker's status information is generated.

[0055] Optionally, the step of constructing a dynamic risk assessment model for high-altitude construction based on the status information of high-altitude workers, quantifying the risks of abnormal postures and safety rope detachment, and generating dynamic risk results for high-altitude construction includes:

[0056] Based on the status information of high-altitude workers, a dynamic risk assessment model for high-altitude construction is constructed, including rules for quantifying abnormal posture risks, rules for quantifying safety rope detachment risks, rules for fusion of composite risks, and rules for mapping risk levels.

[0057] Based on the abnormal posture risk quantification rules, the abnormal posture risk of high-altitude workers is dynamically quantified and analyzed to generate abnormal posture risk values.

[0058] Based on the safety rope detachment risk quantification rules, a dynamic quantitative analysis of the safety rope detachment risk of high-altitude workers is conducted to generate a safety rope detachment risk value.

[0059] Based on the abnormal posture risk value, the safety rope detachment risk value, the human posture change status, and the safety rope connection status, dynamic risk results for high-altitude construction are generated.

[0060] Based on the dynamic risk results of high-altitude construction, the corresponding safety risk level of high-altitude workers is determined, and a high-altitude construction safety inspection response instruction is generated.

[0061] The beneficial effects of this invention are:

[0062] (1) This invention introduces physical safety constraint coefficients and spatial topological relationships into the CoTracker temporal sliding window attention module, and performs dynamic weight modulation when performing cross-frame association of dense point matrix feature representations. This makes the point feature association weight no longer rely solely on visual similarity, but simultaneously integrates spatial constraint information of safe operation boundary areas, safety rope connection constraint areas, and high-altitude fall risk areas. This enables semantic enhancement association modeling of human body joint feature points, safety rope hook points, and safety anchor position feature points, improves the stability of cross-frame association of dense points, and enhances the tracking continuity of key sensing parts.

[0063] (2) This invention improves the CoTracker motion update module by fusing and mapping the enhanced dense point temporal feature vector, point type identifier, spatial topology relationship and physical safety constraint coefficient. Differentiated displacement search range and candidate displacement response mechanism are set for human joint feature points, safety rope hook points and safety anchor position feature points. Structural constraints and safety semantic constraints are introduced in the motion prediction stage, so that the point displacement increment prediction process has type sensitivity and spatial consistency. It effectively suppresses the drift of safety rope hook points and the jitter of safety anchor position feature points. In particular, it significantly improves the trajectory prediction accuracy and stability in local occlusion and complex background. Attached Figure Description

[0064] 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:

[0065] Figure 1 This is a flowchart of a security inspection method based on edge intelligence and dynamic risk assessment proposed in this invention;

[0066] Figure 2 This is a key point trajectory prediction diagram of the traditional method in Embodiment 4 of the security inspection method based on edge intelligence and dynamic risk assessment proposed in this invention;

[0067] Figure 3 This is a key point trajectory prediction diagram of the method proposed in Implementation 4 of the safety inspection method based on edge intelligence and dynamic risk assessment. Detailed Implementation

[0068] Example 1: Reference Figure 1 A security inspection method based on edge intelligence and dynamic risk assessment includes:

[0069] Collect video inspection image sequences of the target high-altitude construction area to generate a high-altitude construction safety model;

[0070] In this embodiment, generating a high-altitude construction safety model includes:

[0071] Collect continuous video frames generated by fixed monitoring equipment or inspection monitoring equipment in the target high-altitude construction area, and generate a video inspection image sequence.

[0072] Based on the construction structure information of the high-altitude construction area, the video inspection image sequence is spatially calibrated to generate the high-altitude construction area calibration.

[0073] Based on the marking of the high-altitude construction area, the information on the activity range of the workers, and the safety protection requirements, determine the safe operation boundary area, the safety rope connection and restraint area, and the high-altitude fall risk area under the spatial location of the high-altitude construction area.

[0074] In Example 1, based on the high-altitude construction area calibration, the construction platform boundary, edge protection boundary, construction passage boundary, and safety rope fixing position in the video inspection image sequence are unified to the same spatial coordinate system to determine the set of spatial positions that workers are allowed to reach.

[0075] Within the set of spatial locations, calculate the spatial geometric features of each location and classify the regions accordingly:

[0076] The high-altitude construction area whose distance from the boundary of the construction platform is not less than the required safe distance from the edge is defined as the safe operation boundary area.

[0077] The spatial location of the high-altitude construction area where the distance between the fixed position of the safety rope and the fixed position is no greater than the allowable connection distance requirement of the safety rope is defined as the safety rope connection constraint area.

[0078] The spatial location of a high-altitude construction area where the construction platform height meets the requirements for fall height and the distance between the platform and the adjacent protective boundary is less than the distance requirement for fall risk edge is defined as a fall risk zone.

[0079] The safe operation boundary area, the safety rope connection constraint area, and the high-altitude fall risk area are superimposed and labeled. When the spatial location of the same high-altitude construction area meets the judgment conditions of multiple areas at the same time, the spatial location of the high-altitude construction area is marked as the corresponding multiple areas. In subsequent safety inspection and judgment, the risk constraint matching is performed in the order of high-altitude fall risk area, safety rope connection constraint area, and safe operation boundary area.

[0080] The safety distance to the edge is determined based on the minimum horizontal buffer distance from the boundary of the construction platform to the permissible working position, with a value range of 0.5-2m; the permissible connection distance of the safety rope is determined based on the remaining connectable distance after deducting the length of the hook structure, the buffer deployment margin, and the swing safety margin from the effective length of the safety rope, with a value range of 1-2m; the fall height is determined based on the vertical distance from the working surface to the fall height reference plane, with a threshold value of not less than 2m; the distance to the edge of the fall risk is determined based on the horizontal distance from the center of gravity of the worker at height to the edge protection boundary, combined with the fall radius and image spatial calibration error, with a value range of 0.5-1.5m.

[0081] The safe operation boundary area, the safety rope connection constraint area, and the high-altitude fall risk area are projected onto the video inspection image frames in the video inspection image sequence to generate a safe operation pixel mask, a safety rope connection pixel mask, and a high-altitude fall pixel mask.

[0082] In Example 1, when the video inspection image sequence originates from a fixed monitoring device, the safe operation boundary area, the safety rope connection constraint area, and the high-altitude fall risk area are converted into two-dimensional pixel areas in each video inspection image frame according to the fixed spatial mapping relationship.

[0083] When the video inspection image sequence comes from the inspection and monitoring equipment, the safe operation boundary area, the safety rope connection constraint area and the high-altitude fall risk area are converted into two-dimensional pixel areas in the corresponding video inspection image frame according to the frame-by-frame spatial mapping relationship corresponding to each video inspection image frame. Based on the two-dimensional pixel areas, a safe operation pixel mask, a safety rope connection pixel mask and a high-altitude fall pixel mask are generated.

[0084] The safety operation boundary area, safety rope connection constraint area, high-altitude fall risk area, safety operation pixel mask, safety rope connection pixel mask and high-altitude fall pixel mask are associated and encapsulated to generate a high-altitude construction safety model.

[0085] Based on the high-altitude construction safety model, the high-altitude workers in the video inspection image sequence are located, and key sensing parts of the high-altitude workers' body structure and safety protection structure are identified to construct a high-altitude worker safety matrix.

[0086] In this embodiment, based on the high-altitude construction safety model, target localization of high-altitude workers in the video inspection image sequence is performed, including:

[0087] Based on the safety operation pixel mask, safety rope connection pixel mask and high-altitude fall pixel mask in the high-altitude construction safety model, spatial constraint screening is performed on the video inspection image frames to determine the target positioning area of ​​high-altitude workers.

[0088] In Example 1, the safety operation pixel mask, the safety rope connection pixel mask, and the high-altitude fall pixel mask are superimposed on each video inspection image frame in the video inspection image sequence. The image areas that overlap with the safety operation pixel mask, the safety rope connection pixel mask, and the high-altitude fall pixel mask, as well as the edge expansion buffer, are retained. The distant view and non-operation background areas that are unrelated to the target high-altitude construction area are excluded to generate the target positioning area for high-altitude workers.

[0089] Within the target positioning area of ​​the high-altitude workers, the high-altitude workers are located, and the high-altitude workers positioning results are generated.

[0090] In Example 1, the target location area of ​​the high-altitude worker is used as the personnel search range. The image content within the target location area of ​​the high-altitude worker is used to extract the personnel candidate region to obtain a set of personnel candidate regions. The personnel candidate regions that meet the human body shape characteristics are selected to obtain the effective personnel candidate regions. The human body contour edge is extracted to obtain the contour range of the high-altitude worker.

[0091] The outer perimeter of the worker's body is determined based on the outline range. The center position of the worker is determined based on the geometric center of the outer perimeter of the human body and the centroid of the outline range.

[0092] The system correlates the outline of the high-altitude worker, the external area of ​​the worker, and the center position of the worker to generate the positioning result of the high-altitude worker.

[0093] Based on the positioning results of high-altitude workers, key sensory parts of the high-altitude workers' body structure are identified to obtain the human joint feature points of the workers.

[0094] In Example 1, within the personnel outline range corresponding to the high-altitude worker positioning results, the human body structure of the high-altitude worker's head, shoulders, elbows, wrists, hips, knees, and ankles is located to obtain the human body joint feature points of the personnel's posture changes.

[0095] Based on the positioning results of high-altitude workers and the pixel mask of safety rope connection in the high-altitude construction safety model, the key sensing parts of the safety protection structure of high-altitude workers are identified to obtain the safety rope hook points.

[0096] In Example 1, near the waist, chest, or back of the worker corresponding to the positioning result of the high-altitude worker, the safety rope end hook, hook closure, and safety belt connector are locally identified by combining the safety rope connection pixel mask to determine the safety rope hook point corresponding to the high-altitude worker.

[0097] Based on the safety rope connection pixel mask and the high-altitude construction area calibration in the high-altitude construction safety model, the key sensing parts of the anchoring structure corresponding to the fixed position of the safety rope are identified to obtain the safety anchoring position feature points.

[0098] In Example 1, the projection position of the safety rope fixing position in the video inspection image frame is determined according to the high-altitude construction area calibration. The anchoring ring, anchoring rod, fixing point or structural connector is locally identified in the safety rope connection pixel mask to obtain the safety anchoring position feature point.

[0099] Based on the spatial relationship between the joint feature points of the human body, the safety rope buckle points, and the safety anchor position feature points, the key sensing parts corresponding to the same high-altitude worker are associated and grouped to construct a safety point matrix for high-altitude workers.

[0100] In Example 1, the positioning results of high-altitude workers are used as the correlation benchmark. The human joint feature points located within the same outline range of the personnel are determined as the human posture point matrix, and the safety rope hook points bound to the positioning results of high-altitude workers are determined as the safety rope connection point matrix.

[0101] Within the safety rope connection pixel mask, candidate safety anchor position feature points are selected through candidate safety anchor position feature points. The spatial connection relationship between the safety rope hook point and each candidate safety anchor position feature point is determined based on the spatial distance between them, whether they are located within the same safety rope connection pixel mask, and whether the line connecting them passes through the safety rope connection pixel mask.

[0102] Pixel sampling is performed on the line connecting the safety rope hook point and the candidate safety anchor position feature point in the video inspection image frame. The proportion of sampled pixels falling into the safety rope connection pixel mask is counted. When the proportion of sampled pixels falling into the safety rope connection pixel mask is not lower than the preset mask coverage ratio threshold, it is determined that the line connecting the safety rope hook point and the candidate safety anchor position feature point passes through the safety rope connection pixel mask.

[0103] The preset mask coverage ratio threshold is determined by subtracting one standard deviation from the average coverage ratio of the sampled pixels of the line connecting the safety rope hook point and the safety anchor position feature point in the confirmed valid connection state sample from the coverage ratio of the pixel mask of the safety rope connection. The value range is 0.7-0.9.

[0104] When there are safe anchorage feature points that satisfy the spatial connection relationship, the safe anchorage feature point with the smallest spatial distance is determined as the safe anchorage point matrix corresponding to the high-altitude worker.

[0105] When there are no safe anchorage location feature points that satisfy the spatial connection relationship, the candidate safe anchorage location feature point with the smallest spatial distance is determined as the safe anchorage point matrix to be verified, and the connection status of the safety rope corresponding to the high-altitude worker is marked as the connection status to be judged.

[0106] The human posture dot matrix, safety rope connection dot matrix, and safety anchor dot matrix or safety anchor dot matrix to be verified are associated and grouped to generate a safety dot matrix for high-altitude workers.

[0107] Based on the safety dot matrix of high-altitude workers, dense feature points are initialized for continuous video frames in the video inspection image sequence. The human body joint feature points, safety rope buckle points and safety anchor position feature points are converted into dense dot matrix feature representations. The dense dot matrix feature representations are then input into the improved CoTracker tracking network to generate initial information of high-altitude workers.

[0108] In this embodiment, based on the safety dot matrix for high-altitude workers, dense feature point initialization is performed on consecutive video frames in the video inspection image sequence, including:

[0109] Based on the safety point matrix for high-altitude workers, determine the human joint feature points, safety rope hook points, and safety anchor position feature points corresponding to the same high-altitude worker, as well as the initial range of the corresponding dense feature points.

[0110] In Example 1, the high-altitude worker safety dot matrix is ​​used as the dense feature point initialization object. The high-altitude worker safety dot matrix is ​​used as a safety semantic dot matrix including human posture dot matrix, safety rope connection dot matrix, and safety anchor dot matrix for dense feature point initialization. The human joint feature points, safety rope buckle points, and safety anchor position feature points are respectively used as the center of the corresponding dense feature point initialization range. Based on the human contour range, the safety rope connection pixel mask, and the local structural region where the safety anchor position feature points are located, the dense feature point initialization range of the human joint feature points, the dense feature point initialization range of the safety rope buckle points, and the dense feature point initialization range of the safety anchor position feature points are determined.

[0111] The pixel region within the preset neighborhood is obtained by taking a circular or square pixel window as the center of the human body joint feature point, safety rope buckle point, or safety anchor position feature point, and cropping it within the pixel mask of the human outline or safety rope connection according to the corresponding initialized neighborhood radius. The initialized neighborhood radius is set according to the differences in target scale, structural stability, and safety semantics of the human body joint feature point, safety rope buckle point, and safety anchor position feature point. The initialized neighborhood radius corresponding to the human body joint feature point is 8-24 pixels, the initialized neighborhood radius corresponding to the safety rope buckle point is 4-12 pixels, and the initialized neighborhood radius corresponding to the safety anchor position feature point is 6-18 pixels.

[0112] Neighborhood sampling is performed within the initial range of dense feature points of human joint feature points, safety rope buckle points, and safety anchor position feature points to generate dense sets of human joint points, safety rope buckle points, and safety anchor points.

[0113] The dense set of human joints, the dense set of safety rope buckles, and the dense set of safety anchors are merged to generate a dense set of high-altitude workers, and a dense lattice feature representation with spatial topological relationships is generated.

[0114] In Example 1, the sets of dense points of human joints, dense points of safety rope hooks, and dense points of safety anchors corresponding to the same high-altitude worker are merged, and dense points with duplicate pixel positions are removed. According to the source of the dense points, they are respectively assigned dense point identifiers of human joints, dense point identifiers of safety rope hooks, and dense point identifiers of safety anchors to generate a set of dense points of high-altitude workers.

[0115] The spatial topological relationships corresponding to the dense point set of high-altitude workers are defined as the human body structure connection relationships between joint feature points, the spatial connection relationships between safety rope hook points and safety anchor position feature points, and the binding relationships between safety rope hook points and high-altitude worker positioning results. The spatial topological relationships are labeled with relationship types, including human body structure connection relationships, safety rope spatial connection relationships, and personnel hook binding relationships. The spatial topological relationships are associated and encapsulated with the dense point set of high-altitude workers to generate a dense point matrix feature representation with spatial topological relationships.

[0116] The dense lattice feature representation with spatial topological relationships is input into the improved CoTracker tracking network to generate spatial visual features of continuous video frames.

[0117] In Example 1, the set of densely populated points of high-altitude workers, the spatial topological relationship, and the set of point type identifiers are used together as the point matrix input information of the improved CoTracker tracking network, and the one-to-one correspondence between the three is preserved. The densely populated points of human joints, safety rope buckles, and safety anchors are not used as indiscriminate two-dimensional point sets as input to the improved CoTracker tracking network. The continuous video frames in the video inspection image sequence are used as the image input information of the improved CoTracker tracking network. The continuous video frames are scale normalized and the inter-frame order is preserved, so that the same high-altitude worker in the continuous video frames is kept in a continuous input state in time sequence.

[0118] By improving the feature encoding module in the CoTracker tracking network, hierarchical visual feature extraction is performed on each video inspection image frame to obtain global spatial features, local spatial features of human joints, local spatial features of safety rope buckles, and local spatial features of safety anchors.

[0119] The global spatial features, the local spatial features of human joints, the local spatial features of safety rope buckles, and the local spatial features of safety anchors are spliced ​​and aligned along the channel dimension according to the pixel positions in the video inspection image frames to generate continuous video frame spatial visual features. An index mapping relationship is then established between each channel region of the continuous video frame spatial visual features and the point type identifier set.

[0120] Based on the spatial visual features of continuous video frames, feature sampling is performed on each dense point in the set of dense points of high-altitude workers to generate initial information of high-altitude workers.

[0121] In Example 1, in the spatial visual features of continuous video frames, the corresponding feature position of each dense point in the set of dense points of high-altitude workers is determined according to the pixel position of each dense point in the video inspection image frame. When the pixel position of the dense point is consistent with the feature grid position in the spatial visual features of continuous video frames, the feature value corresponding to the feature grid position is used as the corresponding spatial visual feature vector. When the pixel position of the dense point is located between adjacent feature grids, the feature values ​​corresponding to the adjacent feature grids are interpolated and fused to obtain the spatial visual feature vector corresponding to the dense point.

[0122] The pixel position of each dense point in the video inspection image frame, the point type identifier corresponding to each dense point, the spatial visual feature vector corresponding to each dense point, and the spatial topological relationship are associated and encapsulated to generate the initial information of the high-altitude operation personnel.

[0123] The initial information of the high-altitude workers is input into the CoTracker temporal sliding module, which performs cross-frame correlation calculation on the dense dot matrix feature representation in continuous video frames and dynamically adjusts the dot feature correlation weights to generate an enhanced temporal dot matrix.

[0124] In this embodiment, the initial information of the high-altitude workers is input into the CoTracker time-series sliding module, including:

[0125] Based on the initial information of the high-altitude workers, and according to the video frame acquisition time of the video inspection image sequence, a time-series sliding window is constructed to input information;

[0126] Based on the input information from the temporal sliding window, the CoTracker temporal sliding module performs visual similarity calculations on dense points in consecutive video frames to generate initial point feature association weights.

[0127] In Example 1, within the input information of the time-series sliding window, dense points in the current video inspection image frame are taken as dense points to be associated, and dense points in other consecutive video frames within the same time-series sliding window are taken as candidate dense points to be associated.

[0128] The initial point feature association weights are obtained by normalizing the inner product similarity between the spatial visual feature vectors corresponding to the dense points to be associated and the spatial visual feature vectors corresponding to the candidate dense points to be associated.

[0129] Based on the safe operation boundary area, the safety rope connection constraint area, and the high-altitude fall risk area, determine the physical safety constraint coefficient corresponding to each dense point in the time-series sliding window input information;

[0130] In Example 1, based on the pixel position of each dense point in the video inspection image frame in the time-series sliding window input information, the dense points are matched with the pixel masks for safe operation, safety rope connection, and high-altitude fall, respectively. The region assignment weighted value is obtained by multiplying the assignment status of the dense points to each region's pixel mask by the dimensionless weight coefficient of the corresponding region.

[0131] When the same dense point falls within multiple region pixel masks, the corresponding weighted values ​​of multiple regions are accumulated. When the dense point does not fall within any region pixel mask, only the basic constraint value is retained. The basic constraint value is added to the region weighted value corresponding to the dense point to obtain the physical security constraint coefficient corresponding to the dense point.

[0132] Among them, the basic constraint value is 1, the dimensionless weight coefficient corresponding to the safe operation boundary area ranges from 0.05 to 0.2, the dimensionless weight coefficient corresponding to the safety rope connection constraint area ranges from 0.2 to 0.5, and the dimensionless weight coefficient corresponding to the high-altitude fall risk area ranges from 0.3 to 0.8.

[0133] Based on the spatial topological relationships and point type identifiers in the initial information of high-altitude workers, determine the topological constraint coefficients between the dense points to be associated and the candidate dense points to be associated;

[0134] In Example 1, based on the point type identifiers corresponding to the dense points to be associated and the candidate dense points to be associated, it is determined that the dense points to be associated and the candidate dense points to originate from dense points of human joints, dense points of safety rope hooks, or dense points of safety anchors, respectively. The source feature points corresponding to the dense points to be associated and the source feature points corresponding to the candidate dense points to be associated are matched with the human structure connection relationship, the safety rope spatial connection relationship, and the personnel hook binding relationship in the spatial topology to determine the relationship type between the dense points to be associated and the candidate dense points.

[0135] When both the dense points to be associated and the candidate dense points to be associated originate from dense points of human joints, and the human joint feature points corresponding to the two belong to the same human structure connection relationship, the relationship type between the two is determined to be a human structure connection relationship.

[0136] When one of the dense points to be associated and the candidate dense points to be associated originates from the dense points of safety rope hooks and the other originates from the dense points of safety anchors, and the safety rope hook points and safety anchor position feature points corresponding to the two belong to the same safety rope spatial connection relationship, the relationship type between the two is determined as a safety rope spatial connection relationship.

[0137] When one of the dense points to be associated and the candidate dense points to be associated originates from the dense points of safety rope hooks and the other originates from the dense points of human joints, and the human joint feature points to which the dense points of human joints belong and the safety rope hook points belong to the same high-altitude worker positioning results, the relationship type between the two is determined to be a personnel hook binding relationship; when neither relationship type is satisfied, the relationship type between the two is determined to be a no-topological association relationship.

[0138] When the dense points to be associated and the candidate dense points to be associated simultaneously satisfy multiple relationship types, the main relationship type is determined in the order of safety rope spatial connection relationship, personnel hook binding relationship, and human body structure connection relationship, and the topological constraint coefficient is determined according to the topological weighting value corresponding to the main relationship type.

[0139] The topology constraint coefficient is obtained by superimposing the basic topology value with the topology weighting value corresponding to the main relation type. The basic topology value is 1; when the main relation type is a human body structure connection relationship, the topology weighting value is 0.1-0.3; when the main relation type is a personnel hook binding relationship, the topology weighting value is 0.2-0.5; when the main relation type is a safety rope spatial connection relationship, the topology weighting value is 0.3-0.6; when the main relation type is no topology association relationship, the topology weighting value is 0.

[0140] The initial point feature association weights are dynamically adjusted based on the physical security constraint coefficients and topological constraint coefficients to generate dynamically adjusted point feature association weights.

[0141] Based on the dynamically adjusted point feature association weights, the spatial visual feature vectors corresponding to candidate associated dense points in the temporal sliding window input information are weighted and fused to generate enhanced dense point temporal feature vectors and obtain an enhanced temporal point matrix.

[0142] In Example 1, the dense points to be associated in the current video inspection image frame are used as the enhancement objects. A set of candidate dense points to be associated is determined within the same temporal sliding window, and the dynamically adjusted point feature association weights and spatial visual feature vectors corresponding to each candidate dense point are obtained. The spatial visual feature vectors of each candidate dense point are weighted and accumulated using their respective point feature association weights to obtain the enhanced temporal feature vector of the dense point to be associated.

[0143] The pixel location, point type identifier, spatial topological relationship, physical security constraint coefficient, and enhanced temporal feature vector of the dense points are associated and encapsulated to obtain the enhanced temporal point information corresponding to the dense points.

[0144] All enhanced temporal point information of the same high-altitude worker in the current video inspection image frame is combined according to the index order in the high-altitude worker density point set to generate an enhanced temporal point matrix.

[0145] The improved CoTracker model is defined as a dense point temporal tracking network that integrates physical safety constraints. Its overall structure includes a continuous video frame feature encoding module, a safety semantic dot matrix embedding module, a CoTracker temporal sliding window attention module integrating physical safety constraints, an improved CoTracker motion update module, and a trajectory output module. The model input includes video inspection image sequences, a set of dense points representing high-altitude workers, point type identifiers, spatial topological relationships, physical safety constraint coefficients, safety operation pixel masks, safety rope connection pixel masks, and high-altitude fall pixel masks. The input dimension of the video inspection image sequence is batch size × number of consecutive video frames × number of image channels × image height × image width, with 3 image channels. The input dimension of the set of dense points representing high-altitude workers is batch size ×... The number of dense points × the dimension of two-dimensional pixel coordinates, with the two-dimensional pixel coordinate dimension being 2, is used in the continuous video frame feature encoding module. This module employs 4 layers of convolutional feature extraction layers and 4 layers of visual Transformer encoding layers to map video inspection image frames into continuous video frame spatial visual features with a channel dimension of 256. The security semantic dot matrix embedding module uses 2 layers of multilayer perceptron to map the pixel positions of dense points, point type identifiers, spatial topological relationships, and physical security constraint coefficients into dot matrix embedding features with a channel dimension of 256. The CoTracker temporal sliding window attention module, which integrates physical security constraints, employs 6 layers of temporal Transformer attention layers to perform cross-frame association calculations on the dense dot matrix feature representations within the temporal sliding window. The point feature association weights are dynamically adjusted using physical security constraint coefficients and topological constraint coefficients.

[0146] In this embodiment, the improved CoTracker motion update module employs a 3-layer multilayer perceptron and a 2-layer cyclic motion update layer. It performs incremental point displacement prediction on the enhanced temporal points of human joints, safety rope hooks, and safety anchors. The model output includes the enhanced temporal point matrix, the trajectory of the high-altitude worker, the trajectory of human joint feature points, the trajectory of the safety rope hook point, and the trajectory of the safety anchor position feature points. The output dimension is batch size × number of consecutive video frames × number of dense points × 2D pixel coordinate dimension. The model parameter scale is 35 million to 45 million, with a schematic scale of 39 million. The training method combines general dense point tracking pre-training with high-altitude construction safety inspection scenario fine-tuning. First, dense point temporal tracking pre-training is performed based on publicly available or self-built continuous video point trajectory samples. Then, scenario fine-tuning is performed based on video inspection image sequences collected by fixed monitoring equipment or inspection monitoring equipment in the high-altitude construction area. The training data sources include high-altitude construction video samples with annotations of human joint feature points, safety rope buckle points, safety anchor position feature points, regional pixel masks, human posture change states, and safety rope connection states. Training enhancement samples are generated through occlusion enhancement, illumination disturbance, viewpoint disturbance, and dense point loss simulation.

[0147] Compared to the existing CoTracker, which only uses general two-dimensional query points and visual temporal features for point trajectory estimation, the improvement of this application lies in introducing the safe operation boundary region, safety rope connection constraint region, and high-altitude fall risk region in the high-altitude construction safety model into the temporal sliding window attention module. The human body joint feature points, safety rope hook points, and safety anchor position feature points are constructed into a safety semantic point matrix with point type identification and spatial topological relationship. In the motion update stage, differentiated displacement search ranges are set according to different point types. At the same time, spatial stable trajectory prediction is performed on the safety anchor position feature points. This enables the model to maintain the continuity of key point trajectories even when scaffolding, personnel limbs, safety ropes, and anchoring structures are partially obscured. It improves the stability of the determination of the relative motion relationship between safety rope hook points and safety anchor position feature points, and provides a trajectory basis with high-altitude construction safety semantic constraints for the subsequent generation of human posture change state, safety rope connection state, and high-altitude construction dynamic risk results.

[0148] Based on the enhanced temporal matrix, the improved CoTracker motion update module is used to continuously predict the trajectory of personnel joint feature points, safety rope hook points and safety anchor position feature points to obtain the movement trajectory of high-altitude workers.

[0149] In this embodiment, based on the enhanced temporal point matrix, the improved CoTracker motion update module is used to perform continuous trajectory prediction on the joint feature points of the human body, the safety rope hook points, and the safety anchor position feature points, including:

[0150] Based on the enhanced timing point matrix, determine the enhanced timing points of human joints, safety rope buckles, and safety anchors for the same high-altitude worker;

[0151] In Example 1, the enhanced temporal point matrix of the same high-altitude worker in the current video inspection image frame is used as the motion prediction object. Based on the index correspondence between the dense points, point type identifiers, spatial topological relationships, physical safety constraint coefficients, and the enhanced dense point temporal feature vectors maintained in the enhanced temporal point matrix, the point type identifier and source feature point corresponding to each enhanced temporal point are determined.

[0152] Enhancement time-series points originating from densely populated human joints are assigned to human joint enhancement time-series points, enhancement time-series points originating from densely populated safety rope buckles are assigned to safety rope buckle enhancement time-series points, and enhancement time-series points originating from densely populated safety anchors are assigned to safety anchor enhancement time-series points.

[0153] Input the enhanced timing points of human joints, safety rope buckles, and safety anchors into the improved CoTracker motion update module to obtain a set of displacement increments of the enhanced timing points.

[0154] In Example 1, the improved CoTracker motion update module uses the pixel position, point type identifier, spatial topology, physical safety constraint coefficient, and enhanced dense point temporal feature vector of the enhanced time-series points as motion update inputs to predict the point displacement increment between adjacent video inspection image frames for the enhanced time-series points of human joints, enhanced time-series points of safety rope buckles, and enhanced time-series points of safety anchors.

[0155] For each enhanced temporal point, the corresponding displacement search range is determined based on the pixel position in the current video inspection image frame, according to the corresponding point type identifier, and multiple candidate displacement positions are generated. A multilayer perceptron is used to stitch together and fuse the enhanced dense point temporal feature vectors, point type identifiers, spatial topological relationships, and physical security constraint coefficients to obtain the candidate displacement response values ​​corresponding to each candidate displacement position.

[0156] The displacement search range is determined based on the point type identifier of the enhanced time sequence points. The displacement search range corresponding to the enhanced time sequence points of human joints is greater than that corresponding to the enhanced time sequence points of safety rope hooks, and the displacement search range corresponding to the enhanced time sequence points of safety rope hooks is greater than that corresponding to the enhanced time sequence points of safety anchors.

[0157] The candidate displacement position with the largest candidate displacement response value is determined as the predicted arrival position of the enhanced temporal point in the next video inspection image frame. The combination of the horizontal and vertical pixel differences between the enhanced temporal point and the current pixel position is determined as the point displacement increment. The point displacement increments corresponding to all human joint enhanced temporal points, safety rope buckle enhanced temporal points, and safety anchor enhanced temporal points are arranged in the order of the enhanced temporal point index to obtain the set of enhanced temporal point displacement increments.

[0158] Based on the enhanced temporal point displacement increment set, the pixel position of each enhanced temporal point in the video inspection image sequence is continuously updated to generate an enhanced prediction trajectory set.

[0159] Based on the enhanced predicted trajectory set and spatial topological relationships, the enhanced continuous predicted trajectories originating from the same person's human joint feature points are aggregated to generate human joint feature point trajectories.

[0160] Based on the enhanced predicted trajectory set and spatial topological relationships, the enhanced continuous predicted trajectories originating from the same safety rope hook point are aggregated to generate the motion trajectory of the safety rope hook point.

[0161] In Example 1, based on the index correspondence between the enhanced time-series points and the source feature points, the enhanced continuous prediction trajectory originating from the same safety rope hook point is determined in the enhanced time-series points of the safety rope hook. The frame-by-frame coordinate arithmetic average of multiple enhanced continuous prediction trajectories corresponding to the same safety rope hook point is aggregated according to the video frame acquisition time to obtain the hook prediction position of the safety rope hook point.

[0162] The predicted positions of the same safety rope hook point in continuous video inspection image frames are connected sequentially according to the video frame acquisition time to generate the movement trajectory of the safety rope hook point.

[0163] Based on the enhanced prediction trajectory set and spatial topological relationships, the enhanced continuous prediction trajectories originating from the same safe anchorage location feature point are spatially stabilized to generate the safe anchorage location feature point trajectory.

[0164] In Example 1, based on the index correspondence between the enhanced time-series points and the source feature points, the enhanced continuous prediction trajectory originating from the same safety anchor position feature point is determined in the enhanced time-series points of safety anchoring. The multiple enhanced continuous prediction trajectories corresponding to the same safety anchor position feature point are aggregated by frame-by-frame coordinate arithmetic average according to the video frame acquisition time to obtain the anchoring candidate position of the safety anchor position feature point in each video inspection image frame.

[0165] Taking the anchorage candidate positions of the same safety anchorage location feature point in continuous video inspection image frames as the processing object, the pixel displacement change between adjacent anchorage candidate positions is calculated.

[0166] When the pixel displacement change between adjacent anchoring candidate positions is not greater than the preset anchoring stability pixel threshold, the anchoring candidate position is retained as the anchoring stability position. When the pixel displacement change between adjacent anchoring candidate positions is greater than the preset anchoring stability pixel threshold, the anchoring candidate position is stabilized and corrected according to the statistical results of the center of all anchoring candidate positions within the time-series sliding window of the same safe anchoring position feature point, and the anchoring stability position is obtained.

[0167] Central statistics are obtained by determining the geometric center of the predicted pixel position of all enhanced temporal points corresponding to the same source feature point in the same video inspection image frame.

[0168] The preset anchoring stability pixel threshold is determined based on the pixel scale of the local structural region where the feature point of the safe anchoring position is located, the spatial calibration error of the video inspection image frame, and the allowable observation jitter of the fixed anchoring structure in the continuous video inspection image frame. The value range is 2-8 pixels.

[0169] The anchoring stability positions of the same safety anchoring location feature points in continuous video inspection image frames are connected sequentially according to the video frame acquisition time to generate the safety anchoring location feature point trajectory.

[0170] The trajectory of human joint feature points, the movement trajectory of safety rope buckle points, and the trajectory of safety anchor position feature points are associated with the same high-altitude worker to generate the movement trajectory of the high-altitude worker.

[0171] Based on the movement trajectory of the high-altitude workers, the dynamic analysis of the posture changes and safety rope connection status of the high-altitude workers in continuous video frames is performed to generate the status information of the high-altitude workers.

[0172] In this embodiment, based on the movement trajectory of the high-altitude worker, dynamic analysis is performed on the worker's posture changes and safety rope connection status in continuous video frames, including:

[0173] Based on the trajectory of human joint feature points, the movement trajectory of safety rope hook points, and the trajectory of safety anchor position feature points, the trajectory is aligned according to the video frame acquisition time to generate a trajectory sequence for analyzing the status of high-altitude workers.

[0174] Based on the trajectory of human joint feature points, calculate the geometric features of human posture of high-altitude workers in continuous video frames, and generate a sequence of posture geometric features.

[0175] In Example 1, based on the trajectory of the human joint feature points of the same high-altitude worker, the predicted joint positions corresponding to the head, shoulders, hips, knees and ankles are determined, and the center position of the human body, the trunk tilt angle, the posture height ratio and the joint movement intensity are calculated.

[0176] The center position of the human body is obtained by taking the arithmetic mean of the coordinates of the predicted joint positions corresponding to the joint feature points of all personnel in the same video inspection image frame.

[0177] The torso tilt angle is obtained by measuring the angle between the direction of the line connecting the predicted positions of the shoulder and hip joints and the longitudinal direction of the image.

[0178] The attitude height ratio is obtained by dividing the vertical pixel span of the predicted positions of all joints by the sum of the horizontal pixel span and the attitude stability correction value.

[0179] The joint motion intensity is obtained by statistically analyzing and arithmetically averaging the changes in the predicted position of the same joint in adjacent image frames. The attitude stability correction value is determined based on the pixel calibration error of the video inspection image frame and a minimal constant to prevent the denominator from being zero, with a value range of 1-10 pixels.

[0180] The body's center position, torso tilt angle, posture height ratio, and joint movement intensity are bound and arranged sequentially according to the video frame acquisition time to generate a posture geometric feature sequence.

[0181] Based on the sequence of posture geometric features, the state of human posture changes of high-altitude workers in continuous video frames is determined, and the human posture change state is generated.

[0182] In Example 1, the state of a worker's posture changes in consecutive video frames is determined based on the torso tilt angle, posture height ratio, and joint movement intensity in the posture geometric feature sequence.

[0183] When the torso tilt angle is not greater than the preset torso tilt angle threshold and the posture height ratio is not less than the preset posture height ratio threshold, the human posture change state of the high-altitude worker is determined as the stable working posture state.

[0184] When the torso tilt angle is greater than the preset torso tilt angle threshold and the posture height ratio is not less than the preset fall height ratio threshold, the human posture change state of the high-altitude worker is determined as the bending posture state.

[0185] When the torso tilt angle is not greater than the preset torso tilt angle threshold, and the posture height ratio is less than the preset posture height ratio threshold but not less than the preset fall height ratio threshold, the human posture change state of the high-altitude worker is determined as the squatting work posture state.

[0186] When the posture height ratio is less than the preset fall height ratio threshold and the joint movement intensity is greater than the preset joint movement intensity threshold, the posture change state of the high-altitude worker is determined as a posture change state.

[0187] When the posture height ratio is less than the preset fall height ratio threshold and the joint movement intensity is not greater than the preset joint movement intensity threshold, and this is maintained in multiple consecutive video inspection image frames, the human posture change state of the high-altitude worker is determined as the fall posture state.

[0188] The preset torso tilt angle threshold is determined based on the upper quantile boundary of the torso tilt angle distribution in the stable working posture samples of high-altitude workers, with a value range of 30-60 degrees; the preset posture height ratio threshold is determined based on the lower quantile boundary of the posture height ratio distribution in the upright working posture samples of high-altitude workers, with a value range of 1.2-2.5; the preset fall height ratio threshold is determined based on the upper quantile boundary of the posture height ratio distribution in the fall posture samples of high-altitude workers, with a value range of 0.3-1; the preset joint motion intensity threshold is determined based on the spatial calibration error of the video inspection image frame, the positioning error of the human body joint feature points, and the joint motion intensity distribution in the normal working motion samples of high-altitude workers, with a value range of 5-30 pixels.

[0189] Based on the motion trajectory of the safety rope hook point and the trajectory of the safety anchor position feature point, calculate the relative motion characteristics between the safety rope hook point and the safety anchor position feature point, and generate a safety rope relative connection feature sequence.

[0190] In Example 1, based on the predicted hook position and anchoring stability position of the same worker at height, the safety rope connection distance, the relative displacement change of the hook, and the change in the hook-anchoring direction are calculated:

[0191] The safety rope connection distance is obtained by calculating the pixel Euclidean distance between the predicted position of the hook and the stable position of the anchor in the same video inspection image frame.

[0192] The relative displacement change of the hook is obtained by calculating the difference between the predicted position change of the hook and the stable position change of the anchor in adjacent video inspection image frames.

[0193] The change in the hook-anchor direction is obtained by calculating the angle difference between the connection directions from the predicted hook position to the stable anchor position in adjacent video inspection image frames.

[0194] Using the same high-altitude worker and the same video frame acquisition time as dual indices, the safety rope connection distance, hook relative displacement change, and hook-anchor direction change calculated under the same video inspection image frame are arranged in the order of video frame acquisition time to generate a safety rope relative connection feature sequence.

[0195] Based on the relative connection feature sequence of the safety rope, determine the connection relationship between the safety rope hook point and the safety anchor position feature point corresponding to the same high-altitude worker, and generate the safety rope connection status.

[0196] In Example 1, the connection status of the safety rope between the safety rope hook point and the safety anchor position feature point is determined based on the safety rope connection distance, the relative displacement change of the hook, and the change in the hook-anchor direction in the relative connection feature sequence of the safety rope:

[0197] When the safety rope connection distance is not greater than the preset safety rope connection distance threshold, the relative displacement change of the hook is not greater than the preset hook relative displacement threshold, and the change in the hook-anchor direction is not greater than the preset hook-anchor direction change threshold, the safety rope connection state is determined to be a valid connection state.

[0198] When the safety rope connection distance is greater than the preset safety rope connection distance threshold, or the relative displacement change of the hook is greater than the preset relative displacement threshold of the hook, or the change in the hook-anchoring direction is greater than the preset hook-anchoring direction change threshold, and this condition is maintained in multiple consecutive video inspection image frames, the safety rope connection state is determined to be a disconnected state.

[0199] When at least one of the following exceeds a preset threshold and fails to meet the condition of maintaining the connection for multiple consecutive frames without being disconnected, or when at least one is within the corresponding threshold boundary, the safety rope connection state is determined as a connection state to be judged.

[0200] The preset safety rope connection distance threshold is determined based on the pixel projection length of the allowable connection distance requirement of the safety rope in the video inspection image frame, the positioning error of the safety rope hook point, and the pixel jitter range of the trajectory of the safety anchor position feature point, with a value range of 80-400 pixels; the preset hook relative displacement threshold is determined based on the distribution of the relative displacement change of the safety rope hook point in the effective connection state sample and the spatial calibration error of the video inspection image frame, with a value range of 5-35 pixels; the preset hook-anchor direction change threshold is determined based on the distribution of the change in the connection direction between the safety rope hook point and the safety anchor position feature point in the effective connection state sample, with a value range of 10-45 degrees.

[0201] Based on the changes in human posture and the connection status of the safety rope, and combined with the trajectory sequence of the high-altitude worker's status analysis, the high-altitude worker's status information is generated.

[0202] In Example 1, the following data are associated and encapsulated for the same high-altitude worker: body posture change status, safety rope connection status, body center position, torso tilt angle, posture height ratio, joint movement intensity, safety rope connection distance, relative displacement change of the hook, change of hook-anchor direction, trajectory of human joint feature points, movement trajectory of safety rope hook point, and trajectory of safety anchor position feature points. These data are then arranged sequentially according to the video frame acquisition time to generate high-altitude worker status information.

[0203] Based on the status information of high-altitude workers, a dynamic risk assessment model for high-altitude construction is constructed to quantify the risks of abnormal posture and safety rope detachment, generate dynamic risk results for high-altitude construction, determine the safety risk level, and generate high-altitude construction safety inspection response instructions to realize safety inspection of high-altitude construction areas.

[0204] In this embodiment, a dynamic risk assessment model for high-altitude construction is constructed based on the status information of high-altitude workers, quantifying the risks of abnormal postures and safety rope detachment, and generating dynamic risk results for high-altitude construction, including:

[0205] Based on the status information of high-altitude workers, a dynamic risk assessment model for high-altitude construction is constructed, including rules for quantifying abnormal posture risks, rules for quantifying safety rope detachment risks, rules for fusion of composite risks, and rules for mapping risk levels.

[0206] In Example 1, the status information of the high-altitude workers corresponding to the same high-altitude workers is used as the risk assessment object. The human body posture change status, human body center position, torso tilt angle, posture height ratio, joint movement intensity, safety rope connection status, safety rope connection distance, relative displacement change of the hook, hook-anchor direction change, human body joint feature point trajectory, safety rope hook point movement trajectory, and safety anchor position feature point trajectory are bound according to the video frame acquisition time to form a high-altitude construction dynamic risk assessment model.

[0207] Based on the abnormal posture risk quantification rules, the abnormal posture risk of high-altitude workers is dynamically quantified and analyzed to generate abnormal posture risk values.

[0208] In Example 1, the basic risk value of abnormal posture is determined based on the changes in human posture. The basic risk value of abnormal posture corresponding to the stable working posture is in the range of 0-0.1, the basic risk value of abnormal posture corresponding to the bending and squatting working posture is in the range of 0.2-0.4, the basic risk value of abnormal posture corresponding to the sudden posture is in the range of 0.6-0.8, and the basic risk value of abnormal posture corresponding to the falling posture is in the range of 0.8-1.

[0209] An abnormal posture deviation correction value is generated based on the degree to which the trunk tilt angle exceeds a preset trunk tilt angle threshold, the degree to which the posture height ratio is lower than a preset posture height ratio threshold or a preset fall height ratio threshold, and the degree to which the joint motion intensity exceeds a preset joint motion intensity threshold. The abnormal posture deviation correction value ranges from 0 to 0.3.

[0210] Based on the number of frames in which the human posture change is maintained across multiple consecutive video inspection image frames, an abnormal posture continuous correction value is generated, with the value ranging from 0 to 0.2.

[0211] The abnormal posture base risk value, abnormal posture deviation correction value, and abnormal posture continuous correction value are accumulated, and the accumulated result is limited to between 0 and 1 to generate the abnormal posture risk value.

[0212] Based on the safety rope detachment risk quantification rules, a dynamic quantitative analysis of the safety rope detachment risk of high-altitude workers is conducted to generate a safety rope detachment risk value.

[0213] In Example 1, the basic risk value of the safety rope connection is determined based on the connection status of the safety rope. The basic risk value of the safety rope connection corresponding to the effective connection status ranges from 0 to 0.1, the basic risk value of the safety rope connection corresponding to the connection status to be determined ranges from 0.3 to 0.6, and the basic risk value of the safety rope connection corresponding to the disconnection status ranges from 0.7 to 1.

[0214] Based on the degree to which the safety rope connection distance exceeds the preset safety rope connection distance threshold, the degree to which the relative displacement change of the hook exceeds the preset hook relative displacement threshold, and the degree to which the change in the hook-anchor direction exceeds the preset hook-anchor direction change threshold, a safety rope connection deviation correction value is generated. The value range of the safety rope connection deviation correction value is 0-0.3.

[0215] Based on the number of frames in which the safety rope connection status is maintained across multiple consecutive video inspection image frames, a safety rope connection continuity correction value is generated, with the value ranging from 0 to 0.2.

[0216] The basic risk value of the safety rope connection, the deviation correction value of the safety rope connection, and the continuous correction value of the safety rope connection are accumulated, and the accumulated result is limited to between 0 and 1 to generate the safety rope detachment risk value.

[0217] Based on the abnormal posture risk value, the safety rope detachment risk value, the human posture change status, and the safety rope connection status, dynamic risk results for high-altitude construction are generated.

[0218] In Example 1, the product of the abnormal posture risk value and the safety rope detachment risk value is used as the composite risk enhancement value. By weighting and fusing the abnormal posture risk value, the safety rope detachment risk value, and the composite risk enhancement value when both exist simultaneously, a comprehensive dynamic risk value is obtained, with a value range of 0-1.

[0219] The preset attitude risk fusion weight ranges from 0.3 to 0.5, the preset safety rope risk fusion weight ranges from 0.3 to 0.5, and the preset composite risk fusion weight ranges from 0.1 to 0.3. The sum of the preset attitude risk fusion weight, the preset safety rope risk fusion weight, and the preset composite risk fusion weight is 1.

[0220] Based on the comprehensive dynamic risk value, abnormal posture risk value, safety rope detachment risk value, human posture change status, and safety rope connection status, the dynamic risk results for high-altitude construction are defined, including:

[0221] Normal inspection risk results: The overall dynamic risk value is less than the preset low risk threshold, the abnormal posture risk value is less than the preset abnormal posture attention threshold, the safety rope detachment risk value is less than the preset safety rope detachment attention threshold, the human body posture change state is a stable working posture state, and the safety rope connection state is an effective connection state.

[0222] Anomaly risk result: The anomaly risk value is greater than or equal to the preset anomaly attention threshold, or the human body posture change state is a sudden posture change state or a falling posture state, and the safety rope connection state is a valid connection state.

[0223] Safety rope detachment risk result: The safety rope detachment risk value is greater than or equal to the preset safety rope detachment concern threshold, or the safety rope connection status is detached connection status, and the human body posture change status is not determined to be a posture change state or a falling posture state.

[0224] Composite high-risk result: The human body's posture change state is a sudden posture change state or a fallen posture state, and the safety rope connection state is an undetermined connection state or a disconnected connection state, or the comprehensive dynamic risk value is greater than or equal to the preset high-risk threshold.

[0225] The preset low-risk threshold is obtained by taking the upper quantile boundary of the comprehensive dynamic risk value corresponding to historical normal operation samples, with a value range of 0.2-0.35. The preset abnormal posture concern threshold is obtained by taking the lower quantile boundary of the abnormal posture risk value corresponding to historical abnormal posture samples, with a value range of 0.4-0.6. The preset safety rope detachment concern threshold is obtained by taking the lower quantile boundary of the safety rope detachment risk value corresponding to historical safety rope detachment samples, with a value range of 0.4-0.6. The preset high-risk threshold is obtained by taking the lower quantile boundary of the comprehensive dynamic risk value corresponding to historical composite high-risk samples, with a value range of 0.7-0.9.

[0226] Based on the dynamic risk results of high-altitude construction, the corresponding safety risk level of high-altitude workers is determined, and a high-altitude construction safety inspection response instruction is generated.

[0227] In Example 1, when the dynamic risk result of high-altitude construction is the normal inspection risk result, the safety risk level corresponding to the high-altitude worker is determined to be low risk level, and a routine inspection and maintenance instruction is generated.

[0228] When the dynamic risk result of high-altitude construction is an abnormal posture risk result or a safety rope detachment risk result, and the comprehensive dynamic risk value is less than the preset high risk threshold, the safety risk level corresponding to the high-altitude worker is determined to be a medium risk level, and an abnormal state review instruction is generated.

[0229] When the dynamic risk result of high-altitude construction is an abnormal posture risk result or a safety rope detachment risk result, and the comprehensive dynamic risk value is greater than or equal to the preset high-risk threshold, the safety risk level corresponding to the high-altitude worker is determined as a high-risk level, and a high-risk warning instruction is generated.

[0230] When the dynamic risk result of high-altitude construction is a composite high-risk result, the safety risk level corresponding to the high-altitude workers is determined as an emergency risk level, and an emergency safety handling prompt instruction is generated.

[0231] The dynamic risk results, safety risk level, and safety inspection response instructions for high-altitude construction corresponding to the same high-altitude worker are associated and packaged according to the video frame acquisition time to form the safety inspection results of the high-altitude construction area, thereby realizing the safety inspection of the high-altitude construction area.

[0232] Example 2: A safety inspection system for high-altitude construction areas based on CoTracker and dynamic risk assessment, comprising:

[0233] The acquisition module is used to acquire video inspection image sequences of the target high-altitude construction area and generate a high-altitude construction safety model.

[0234] The positioning module is used to locate workers at height and identify key sensing points to construct a safety dot matrix for workers at height.

[0235] The initialization module is used to initialize dense feature points in consecutive video frames, generate dense dot matrix feature representations and initial information of high-altitude workers;

[0236] The association module is used to perform cross-frame association calculations on the dense dot matrix feature representation, dynamically adjust the dot feature association weights, and generate an enhanced temporal dot matrix.

[0237] The prediction module is used to predict the continuous trajectory of human joint feature points, safety rope hook points and safety anchor position feature points based on the enhanced temporal matrix, so as to obtain the movement trajectory of high-altitude workers.

[0238] The status module is used to analyze the attitude change status and safety rope connection status to generate status information of high-altitude workers;

[0239] The response module is used to build a dynamic risk assessment model for high-altitude construction, and generate dynamic risk results, safety risk levels, and safety inspection response instructions for high-altitude construction.

[0240] Example 3: This example provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0241] The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), registers, hard disks, optical fibers, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof, or any other form of computer-readable storage medium in the art.

[0242] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can reside within an application-specific integrated circuit (ASIC). In embodiments of the invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device.

[0243] Example 4: In this example, within a high-altitude steel structure construction area, the edges of the construction platform, edge guardrails, safety rope anchoring components, scaffolding uprights, horizontal beams, and partial obstruction devices coexist. Workers need to perform movements such as moving, bending, squatting, and tool carrying near the platform edge. Fixed video inspection equipment and edge computing terminals are deployed in the area. The video inspection equipment continuously collects video inspection image sequences with a resolution of 1920×1080 and a frame rate of 25 frames per second. The edge computing terminal performs real-time analysis of the continuous video frames. Traditional inspection methods mainly rely on personnel detection frames, single-frame human posture estimation, and static distance threshold judgment of safety rope hook points. When workers are briefly obstructed by scaffolding uprights, the safety rope is similar in color to the background steel cable, or the anchoring structure is obstructed by horizontal beams, problems such as drift of personnel joint feature points, loss of safety rope hook points, jitter of safety anchoring position feature points, and misjudgment of risk status can easily occur.

[0244] In this embodiment, during the inspection process, the edge computing terminal obtains a continuous video inspection image sequence, which contains 450 frames and covers an 18-second continuous operation. The system generates a high-altitude construction safety model based on the construction platform boundary, edge protection boundary, construction passage boundary, and safety rope fixing position in the video inspection image frames. In the generated result, the pixel mask area of ​​the safe operation boundary area in the image is 412,000 pixels, the pixel mask area of ​​the safety rope connection is 126,500 pixels, the pixel mask area of ​​the fall risk is 86,500 pixels, the pixel projection length corresponding to the edge safety distance is 92 pixels, and the pixel projection length of the allowable connection distance requirement for the safety rope in the video inspection image frame is 236 pixels. Traditional methods only set a single edge warning line in the image, which cannot distinguish between the safe operation boundary area, the safety rope connection constraint area, and the fall risk area. Therefore, false alarms are easily generated when personnel approach but do not enter the fall risk area.

[0245] The system located a worker at height in frames 1 through 40. The worker's outer region was 138 pixels wide and 286 pixels high, with the worker's center located at approximately 860 pixels horizontally and 522 pixels vertically. Within the worker's outline, the system identified 17 human joint feature points corresponding to the head, shoulders, elbows, wrists, hips, knees, and ankles. Simultaneously, it identified a safety rope hook point near the safety harness connection point on the chest and back, and a safety anchor point feature point within the safety rope connection pixel mask. At this point, the pixel Euclidean distance between the safety rope hook point and the safety anchor point feature point was 178 pixels, and the proportion of the line connecting them falling within the safety rope connection pixel mask was 0.84, exceeding the preset mask coverage threshold of 0.78. Therefore, the two were associated as the same safety rope connection point matrix and safety anchor point matrix for the same worker at height. Traditional methods can only reliably identify 29 frames out of 40 in the same video clip because the grayscale of the safety rope hook point is similar to that of the background steel cable, resulting in 11 missed frames.

[0246] During the initialization of dense points, the system generates a dense point matrix feature representation centered on personnel joint feature points, safety rope hook points, and safety anchor position feature points. An average of 9 dense points are generated around each personnel joint feature point, totaling 153 personnel joint dense points; 16 safety rope hook dense points are generated around each safety rope hook point; and 12 safety anchor dense points are generated around each safety anchor position feature point. The set of dense points for the same high-altitude worker comprises 181 dense points. The system assigns dense points with identifiers for personnel joint dense points, safety rope hook dense points, and safety anchor dense points, respectively, and writes the human body structure connection relationship, safety rope spatial connection relationship, and personnel hook binding relationship into the spatial topology. Traditional methods only track key personnel joints as independent points, without distinguishing between safety rope hook points and safety anchor position feature points, and without establishing a safety rope spatial connection relationship. Therefore, in subsequent frames, it is impossible to reliably express whether the safety rope maintains an effective connection with the anchoring structure.

[0247] Between frames 96 and 145, the worker moves from the center of the platform to near the edge, briefly obscured by the scaffolding poles. At this point, the traditional CoTracker model shows wrist keypoint drift in frame 118 after the obstruction, with the error between the wrist keypoint and the actual labeled position increasing from 5.9 pixels to 23.4 pixels. The safety rope hook point experiences trajectory interruption between frames 121 and 130. The method in this embodiment inputs the initial information of the worker at height into the CoTracker temporal sliding window attention module, which integrates physical safety constraints. During cross-frame association, physical safety constraint coefficients are generated based on whether dense points fall within the safety operation pixel mask, the safety rope connection pixel mask, and the fall-from-height pixel mask. The physical safety constraint coefficient for dense points of safety rope hooks located within the safety rope connection pixel mask increases from 1.00 to 1.38; for dense points simultaneously satisfying both the fall-from-height pixel mask and the safety rope connection pixel mask, the physical safety constraint coefficient increases to 1.72. After dynamic adjustment, the point feature association weight between dense points of safety rope hooks and candidate points under the same spatial connection relationship of safety ropes increased from 0.41 to 0.68, while the point feature association weight of candidate points of background steel cables decreased from 0.36 to 0.17, maintaining the continuity of the trajectory of safety rope hook points within the obstruction area.

[0248] During motion updates, the system employs differentiated displacement search ranges for enhanced timing points of human joints, safety rope hooks, and safety anchors. The displacement search range for enhanced timing points of human joints is set to 32 pixels, for safety rope hooks to 18 pixels, and for safety anchors to 8 pixels. In frame 126, 81 candidate displacement positions for the safety rope hooks are generated, with the maximum candidate displacement response value being 0.87, corresponding to a predicted arrival position of 934 pixels horizontally and 471 pixels vertically. The maximum response candidate position given by the traditional method is located at the background steel cable, with 951 pixels horizontally and 454 pixels vertically, resulting in an error of 27.2 pixels compared to the manually labeled hook position. The error of the method in this embodiment is 4.6 pixels. For the safety anchor position feature points, the system uses spatial stabilization processing to control the frame-by-frame jitter of the anchor candidate positions to within 2.3 pixels, while the traditional method experiences anchor point jitter of up to 9.8 pixels when obscured by a beam.

[0249] Between frames 180 and 260, the worker bends over to lift objects. The system calculates the center position of the human body, the torso tilt angle, the posture height ratio, and the joint motion intensity based on the trajectory of human joint feature points. At frame 190, the torso tilt angle is 28 degrees, the posture height ratio is 1.74, and the joint motion intensity is 6.2 pixels, which the system determines to be a stable working posture. At frame 214, the torso tilt angle increases to 47 degrees, the posture height ratio is 1.31, and the joint motion intensity is 9.4 pixels, which the system determines to be a bending working posture. At frame 238, the torso tilt angle is 51 degrees, the posture height ratio is 1.24, and the joint motion intensity is 10.1 pixels, and the system continues to maintain the bending working posture. The traditional single-frame pose method miscalculated the pose height ratio as 0.88 in frame 214 because the shoulder key point was obscured by the safety rope, triggering a false fall alarm. The method in this embodiment maintains the consistency of the shoulder, hip and knee trajectories through continuous trajectory and spatial topology, and does not trigger the risk of fall.

[0250] Between frames 286 and 340, as the worker moves towards the edge, the movement of the safety rope hook point relative to the safety anchor point gradually increases. In frame 286, the safety rope connection distance is 184 pixels, the relative displacement change of the hook is 7.6 pixels, and the change in the hook-anchor direction is 12 degrees, all below the preset threshold. The system determines the safety rope connection status as a valid connection. In frame 312, the safety rope connection distance increases to 262 pixels, the relative displacement change of the hook increases to 21.3 pixels, and the change in the hook-anchor direction is 28 degrees. The safety rope connection distance is within the preset safety rope connection distance threshold of 80-400 pixels but close to the upper limit of the scene's valid connection. The system temporarily sets the safety rope connection status as a pending connection status. From frames 327 to 336, the safety rope connection distance continuously exceeded 318 pixels, the relative displacement change of the hook continuously exceeded 36 pixels, and the change in the hook-anchor direction continuously exceeded 46 degrees. These exceed the preset relative hook displacement thresholds of 5-35 pixels and the preset hook-anchor direction change thresholds of 10-45 degrees, and meet the condition of being maintained for multiple consecutive frames. Therefore, the system determines the connection to be lost. Traditional methods only judge based on the distance between the hook and anchor point in a single frame, outputting a disconnection alarm as early as frame 312, causing premature false alarms. However, from frames 327 to 336, due to brief anchor point jitter, the traditional method actually did not continuously issue alarms for 4 frames.

[0251] During the risk assessment phase, the system constructs a dynamic risk assessment model for high-altitude construction based on the status information of the high-altitude workers. In frame 238, the human posture change is a bent-over working posture, the safety rope connection is a valid connection, the basic risk value for abnormal posture is 0.32, the deviation correction value for abnormal posture is 0.07, the continuous correction value for abnormal posture is 0.05, and the overall risk value for abnormal posture is 0.44; the basic risk value for safety rope connection is 0.06, the deviation correction value for safety rope connection is 0.04, the continuous correction value for safety rope connection is 0.00, and the risk value for safety rope detachment is 0.10; the overall dynamic risk value is 0.31. The system generates an abnormal posture risk result, determines the safety risk level to be medium risk, and generates an abnormal status review instruction. In frame 332, the human body's posture change is a bent-over working posture, and the safety rope connection status is detached. The abnormal posture risk value is 0.46, the safety rope detachment risk value is 0.82, the composite risk enhancement value is 0.38, and the comprehensive dynamic risk value is 0.66. The system generates a safety rope detachment risk result, determines the safety risk level as medium risk, and generates an abnormal state review instruction. In frame 354, the worker experiences a sudden posture change near the edge, with the posture height ratio decreasing to 0.58 and the joint movement intensity increasing to 27.5 pixels. Simultaneously, the safety rope connection status remains detached. The abnormal posture risk value is 0.79, the safety rope detachment risk value is 0.86, the composite risk enhancement value is 0.68, and the comprehensive dynamic risk value is 0.82. The system generates a composite high-risk result, determines the safety risk level as emergency risk, and generates an emergency safety handling prompt instruction. This includes the target personnel number, video frame range, human posture change status, safety rope connection status, comprehensive dynamic risk value, and review screen segment index, and is sent to the edge safety inspection terminal.

[0252] To verify feasibility, this embodiment constructed training and testing samples for high-altitude construction inspection. The training samples contain 960 video inspection image sequences, with a total of 241,600 frames. Among them, there are 310 samples of stable working posture, 185 samples of bending working posture, 142 samples of squatting working posture, 96 samples of posture change, 74 samples of falling posture, 492 samples of effective connection, 178 samples of connection to be determined, and 290 samples of disconnection. Each sample is labeled with human joint feature points, safety rope hook points, safety anchor position feature points, safety operation pixel mask, safety rope connection pixel mask, high-altitude fall pixel mask, human posture change state, and safety rope connection state. The test sample contains 240 video inspection image sequences, with a total of 60,400 frames in the video clips. Among them are 76 samples with slight occlusion, 91 samples with moderate occlusion, 42 samples with severe occlusion, 18 samples with strong light reflection, and 13 samples with background steel cable interference.

[0253] Several representative segments were selected from the training samples for verification. In sample segment A, the worker moves upright and passes through the obstruction of the scaffold pole. The obstruction lasts for 17 frames. Manual annotation shows that the average displacement of the human joints is 5.1 pixels per frame, the average displacement of the safety rope hook point is 3.4 pixels per frame, and the actual displacement of the safety anchor position feature point is less than 1 pixel. The traditional method shows two trajectory breaks within the obstruction area, while the method in this embodiment does not show trajectory breaks. In sample segment B, the worker bends over to carry materials and overlaps with the safety rope. The traditional method mismatches the safety rope hook point to the background steel cable in 18.6% of cases, while the method in this embodiment reduces this to 3.2%. In sample segment C, the safety rope hook point gradually moves away from the safety anchor position feature point and enters a disengaged state. The traditional method has an accuracy rate of 79.4% in identifying the boundary between the connection state and the disengaged state, while the method in this embodiment achieves 92.8%. In sample segment D, the worker experiences a sudden change in posture accompanied by an abnormal connection of the safety rope. Traditional methods only output the risk result of the abnormal posture in the first 15 frames. The method in this embodiment improves the risk result to a composite high-risk result in 6 consecutive frames.

[0254] In the test sample comparison, the traditional method uses general CoTracker trajectory tracking, single-frame pose estimation, and fixed distance threshold judgment. The method in this embodiment uses a CoTracker temporal sliding window attention module that integrates physical safety constraints, an improved CoTracker motion update module, and a dynamic risk assessment model for high-altitude construction. Regarding the average tracking error of personnel joint feature points, the traditional method is 7.4 pixels, while the method in this embodiment is 3.6 pixels; regarding the average tracking error of safety rope hook points, the traditional method is 10.8 pixels, while the method in this embodiment is 4.1 pixels; regarding the spatial stability error of safety anchor position feature points, the traditional method is 6.2 pixels, while the method in this embodiment is 2.0 pixels; regarding the number of frames required for occlusion recovery, the traditional method requires an average of 11.7 frames to recover a stable trajectory, while the method in this embodiment requires an average of 4.3 frames; regarding the trajectory interruption rate, the traditional method is 14.8%, while the method in this embodiment is... The accuracy rate of the method in this embodiment is 3.9%; regarding the accuracy rate of safety rope connection status recognition, the traditional method is 83.1%, while the method in this embodiment is 94.2%; regarding the accuracy rate of human posture change status recognition, the traditional method is 85.6%, while the method in this embodiment is 93.5%; regarding the accuracy rate of composite high-risk result recognition, the traditional method is 76.9%, while the method in this embodiment is 91.8%; regarding the false alarm rate, the traditional method is 12.7%, while the method in this embodiment is 4.6%; regarding the missed alarm rate, the traditional method is 9.5%, while the method in this embodiment is 3.1%.

[0255] In edge computing performance verification, when the input resolution of a single-channel video inspection image sequence is 1920×1080, the traditional method has an average single-frame processing time of 38.5 milliseconds. The method in this embodiment, due to the addition of physical safety constraint calculation and dynamic risk assessment, has an average single-frame processing time of 43.2 milliseconds, but still meets the real-time inspection requirement of 25 frames / second. For a continuous video segment of 450 frames, the traditional method has an average delay of 1.46 seconds in outputting risk results, while the method in this embodiment has an average delay of 0.72 seconds in outputting dynamic risk results for high-altitude construction on the edge computing terminal. This is because the method in this embodiment reduces the number of repeated misjudgments, re-identifications, and manual verification triggers through continuous trajectory and risk maintenance mechanisms. For 52 test segments containing the risk of safety rope detachment, the traditional method correctly triggered alarms in 43 segments and missed 9; the method in this embodiment correctly triggered alarms in 50 segments and missed 2. For 67 test segments containing normal bending or squatting movements without real risk, the traditional method falsely reported 14 segments, while the method in this embodiment falsely reported 4.

[0256] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A security inspection method based on edge intelligence and dynamic risk assessment, characterized in that, include: Collect video inspection image sequences of the target high-altitude construction area to generate a high-altitude construction safety model; Based on the high-altitude construction safety model, the high-altitude workers in the video inspection image sequence are located, and key sensing parts of the high-altitude workers' body structure and safety protection structure are identified to construct a high-altitude worker safety matrix. Based on the safety dot matrix of high-altitude workers, dense feature points are initialized for continuous video frames in the video inspection image sequence. The human body joint feature points, safety rope buckle points and safety anchor position feature points are converted into dense dot matrix feature representations. The dense dot matrix feature representations are then input into the improved CoTracker tracking network to generate initial information of high-altitude workers. The initial information of the high-altitude workers is input into the CoTracker temporal sliding module, which performs cross-frame correlation calculation on the dense dot matrix feature representation in continuous video frames and dynamically adjusts the dot feature correlation weights to generate an enhanced temporal dot matrix. Based on the enhanced temporal matrix, the improved CoTracker motion update module is used to continuously predict the trajectory of personnel joint feature points, safety rope hook points and safety anchor position feature points to obtain the movement trajectory of high-altitude workers. Based on the movement trajectory of the high-altitude workers, the dynamic analysis of the posture changes and safety rope connection status of the high-altitude workers in continuous video frames is performed to generate the status information of the high-altitude workers. Based on the status information of high-altitude workers, a dynamic risk assessment model for high-altitude construction is constructed to quantify the risks of abnormal posture and safety rope detachment, generate dynamic risk results for high-altitude construction, determine the safety risk level, and generate high-altitude construction safety inspection response instructions to realize safety inspection of high-altitude construction areas.

2. The security inspection method based on edge intelligence and dynamic risk assessment according to claim 1, characterized in that, The generation of the high-altitude construction safety model includes: Collect continuous video frames generated by fixed monitoring equipment or inspection monitoring equipment in the target high-altitude construction area, and generate a video inspection image sequence. Based on the construction structure information of the high-altitude construction area, the video inspection image sequence is spatially calibrated to generate the high-altitude construction area calibration. Based on the marking of the high-altitude construction area, the information on the activity range of the workers, and the safety protection requirements, determine the safe operation boundary area, the safety rope connection and restraint area, and the high-altitude fall risk area under the spatial location of the high-altitude construction area. The safe operation boundary area, the safety rope connection constraint area, and the high-altitude fall risk area are projected onto the video inspection image frames in the video inspection image sequence to generate a safe operation pixel mask, a safety rope connection pixel mask, and a high-altitude fall pixel mask. The safety operation boundary area, safety rope connection constraint area, high-altitude fall risk area, safety operation pixel mask, safety rope connection pixel mask and high-altitude fall pixel mask are associated and encapsulated to generate a high-altitude construction safety model.

3. The security inspection method based on edge intelligence and dynamic risk assessment according to claim 2, characterized in that, The step of locating high-altitude workers in a video inspection image sequence based on a high-altitude construction safety model includes: Based on the safety operation pixel mask, safety rope connection pixel mask and high-altitude fall pixel mask in the high-altitude construction safety model, spatial constraint screening is performed on the video inspection image frames to determine the target positioning area of ​​high-altitude workers. Within the target positioning area of ​​the high-altitude workers, the high-altitude workers are located, and the high-altitude workers positioning results are generated. Based on the positioning results of high-altitude workers, key sensory parts of the high-altitude workers' body structure are identified to obtain the human joint feature points of the workers. Based on the positioning results of high-altitude workers and the pixel mask of safety rope connection in the high-altitude construction safety model, the key sensing parts of the safety protection structure of high-altitude workers are identified to obtain the safety rope hook points. Based on the safety rope connection pixel mask and the high-altitude construction area calibration in the high-altitude construction safety model, the key sensing parts of the anchoring structure corresponding to the fixed position of the safety rope are identified to obtain the safety anchoring position feature points. Based on the spatial relationship between the joint feature points of the human body, the safety rope buckle points, and the safety anchor position feature points, the key sensing parts corresponding to the same high-altitude worker are associated and grouped to construct a safety point matrix for high-altitude workers.

4. The security inspection method based on edge intelligence and dynamic risk assessment according to claim 3, characterized in that, The initialization of dense feature points for consecutive video frames in the video inspection image sequence based on the safety dot matrix for high-altitude workers includes: Based on the safety point matrix for high-altitude workers, determine the human joint feature points, safety rope hook points, and safety anchor position feature points corresponding to the same high-altitude worker, as well as the initial range of the corresponding dense feature points. Neighborhood sampling is performed within the initial range of dense feature points of human joint feature points, safety rope buckle points, and safety anchor position feature points to generate dense sets of human joint points, safety rope buckle points, and safety anchor points. The dense set of human joints, the dense set of safety rope buckles, and the dense set of safety anchors are merged to generate a dense set of high-altitude workers, and a dense lattice feature representation with spatial topological relationships is generated. The dense lattice feature representation with spatial topological relationships is input into the improved CoTracker tracking network to generate spatial visual features of continuous video frames. Based on the spatial visual features of continuous video frames, feature sampling is performed on each dense point in the set of dense points of high-altitude workers to generate initial information of high-altitude workers.

5. A security inspection method based on edge intelligence and dynamic risk assessment according to claim 4, characterized in that, The process of inputting the initial information of high-altitude workers into the CoTracker time-series sliding module includes: Based on the initial information of the high-altitude workers, and according to the video frame acquisition time of the video inspection image sequence, a time-series sliding window is constructed to input information; Based on the input information from the temporal sliding window, the CoTracker temporal sliding module performs visual similarity calculations on dense points in consecutive video frames to generate initial point feature association weights. Based on the safe operation boundary area, the safety rope connection constraint area, and the high-altitude fall risk area, determine the physical safety constraint coefficient corresponding to each dense point in the time-series sliding window input information; Based on the spatial topological relationships and point type identifiers in the initial information of high-altitude workers, determine the topological constraint coefficients between the dense points to be associated and the candidate dense points to be associated; The initial point feature association weights are dynamically adjusted based on the physical security constraint coefficients and topological constraint coefficients to generate dynamically adjusted point feature association weights. Based on the dynamically adjusted point feature association weights, the spatial visual feature vectors corresponding to candidate associated dense points in the temporal sliding window input information are weighted and fused to generate enhanced dense point temporal feature vectors and obtain an enhanced temporal point matrix.

6. A security inspection method based on edge intelligence and dynamic risk assessment according to claim 5, characterized in that, The step of predicting the continuous trajectory of human joint feature points, safety rope buckle points, and safety anchor position feature points using the improved CoTracker motion update module based on the enhanced temporal point matrix includes: Based on the enhanced timing point matrix, determine the enhanced timing points of human joints, safety rope buckles, and safety anchors for the same high-altitude worker; Input the enhanced timing points of human joints, safety rope buckles, and safety anchors into the improved CoTracker motion update module to obtain a set of displacement increments of the enhanced timing points. Based on the enhanced temporal point displacement increment set, the pixel position of each enhanced temporal point in the video inspection image sequence is continuously updated to generate an enhanced prediction trajectory set. Based on the enhanced predicted trajectory set and spatial topological relationships, the enhanced continuous predicted trajectories originating from the same person's human joint feature points are aggregated to generate human joint feature point trajectories. Based on the enhanced predicted trajectory set and spatial topological relationships, the enhanced continuous predicted trajectories originating from the same safety rope hook point are aggregated to generate the motion trajectory of the safety rope hook point. Based on the enhanced prediction trajectory set and spatial topological relationships, the enhanced continuous prediction trajectories originating from the same safe anchorage location feature point are spatially stabilized to generate the safe anchorage location feature point trajectory. The trajectory of human joint feature points, the movement trajectory of safety rope buckle points, and the trajectory of safety anchor position feature points are associated with the same high-altitude worker to generate the movement trajectory of the high-altitude worker.

7. A security inspection method based on edge intelligence and dynamic risk assessment according to claim 6, characterized in that, The dynamic analysis of the posture changes and safety rope connection status of the high-altitude workers in continuous video frames based on their movement trajectory includes: Based on the trajectory of human joint feature points, the movement trajectory of safety rope hook points, and the trajectory of safety anchor position feature points, the trajectory is aligned according to the video frame acquisition time to generate a trajectory sequence for analyzing the status of high-altitude workers. Based on the trajectory of human joint feature points, calculate the geometric features of human posture of high-altitude workers in continuous video frames, and generate a sequence of posture geometric features. Based on the sequence of posture geometric features, the state of human posture changes of high-altitude workers in continuous video frames is determined, and the human posture change state is generated. Based on the motion trajectory of the safety rope hook point and the trajectory of the safety anchor position feature point, calculate the relative motion characteristics between the safety rope hook point and the safety anchor position feature point, and generate a safety rope relative connection feature sequence. Based on the relative connection feature sequence of the safety rope, determine the connection relationship between the safety rope hook point and the safety anchor position feature point corresponding to the same high-altitude worker, and generate the safety rope connection status. Based on the changes in human posture and the connection status of the safety rope, and combined with the trajectory sequence of the high-altitude worker's status analysis, the high-altitude worker's status information is generated.

8. A security inspection method based on edge intelligence and dynamic risk assessment according to claim 7, characterized in that, The process involves constructing a dynamic risk assessment model for high-altitude construction based on the status information of high-altitude workers, quantifying the risks of abnormal postures and safety rope detachment, and generating dynamic risk results for high-altitude construction, including: Based on the status information of high-altitude workers, a dynamic risk assessment model for high-altitude construction is constructed, including rules for quantifying abnormal posture risks, rules for quantifying safety rope detachment risks, rules for fusion of composite risks, and rules for mapping risk levels. Based on the abnormal posture risk quantification rules, the abnormal posture risk of high-altitude workers is dynamically quantified and analyzed to generate abnormal posture risk values. Based on the safety rope detachment risk quantification rules, a dynamic quantitative analysis of the safety rope detachment risk of high-altitude workers is conducted to generate a safety rope detachment risk value. Based on the abnormal posture risk value, the safety rope detachment risk value, the human posture change status, and the safety rope connection status, dynamic risk results for high-altitude construction are generated. Based on the dynamic risk results of high-altitude construction, the corresponding safety risk level of high-altitude workers is determined, and a high-altitude construction safety inspection response instruction is generated.

9. A safety inspection system for high-altitude construction areas based on CoTracker and dynamic risk assessment, used to execute the safety inspection method for high-altitude construction areas based on CoTracker and dynamic risk assessment as described in any one of claims 1-8, characterized in that, include: The acquisition module is used to acquire video inspection image sequences of the target high-altitude construction area and generate a high-altitude construction safety model. The positioning module is used to locate workers at height and identify key sensing points to construct a safety dot matrix for workers at height. The initialization module is used to initialize dense feature points in consecutive video frames, generate dense dot matrix feature representations and initial information of high-altitude workers; The association module is used to perform cross-frame association calculations on the dense dot matrix feature representation, dynamically adjust the dot feature association weights, and generate an enhanced temporal dot matrix. The prediction module is used to predict the continuous trajectory of human joint feature points, safety rope hook points and safety anchor position feature points based on the enhanced temporal matrix, so as to obtain the movement trajectory of high-altitude workers. The status module is used to analyze attitude change status and safety rope connection status to generate status information of high-altitude workers; The response module is used to build a dynamic risk assessment model for high-altitude construction, and generate dynamic risk results, safety risk levels, and safety inspection response instructions for high-altitude construction.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1-8.

Images