A method for identifying high-altitude worker violation behavior based on posture capture

Abstract

The application discloses a kind of high-altitude operation personnel rule violation identification method based on gesture capture, including the following steps: generating high-altitude operation image frame sequence;Scene reference reconstruction is carried out, and high-altitude operation scene reference data is generated;Personnel region anchoring is generated by carrying out human contour segmentation;Personnel gesture observation sequence is generated;Human posture capture is carried out, and operation personnel gesture joint sequence is generated;Gesture trajectory condenses, and personnel gesture trajectory data is generated;Coupling data of gesture boundary is obtained by gesture-boundary coupling projection;Using improved CleverCatch model, candidate rule violation data is generated;Temporal continuity verification and gesture stability check are carried out, and effective rule violation data is generated, and high-altitude operation personnel rule violation identification result is obtained, improve the accuracy and stability of high-altitude operation personnel rule violation identification, reduce artificial patrol omission and ordinary video identification false alarm.

Description

Technical Field

[0001] This invention relates to the field of high-altitude operation safety monitoring technology, and in particular to a method for identifying violations by high-altitude workers based on attitude capture. Background Technology

[0002] With the increasing prevalence of high-altitude work scenarios such as building construction, power maintenance, bridge maintenance, curtain wall installation, and tower work, the safety supervision of high-altitude workers has become a crucial aspect of on-site management. High-altitude work areas typically include work platforms, guardrails, edge areas, areas where personnel can move, and various equipment components. When workers move, bend, reach, climb, cross, or lean out within these areas, the relationship between their body posture and the boundaries of the work space directly impacts work safety. If workers overstep boundaries, illegally cross guardrails, stand abnormally near edges, or shift their center of gravity towards the edge, it can easily lead to accidents such as falls from heights, loss of balance, and collisions. Therefore, accurately capturing changes in the posture of high-altitude workers and identifying violations in conjunction with the spatial boundaries of the high-altitude work area is a key issue that needs to be addressed in the field of high-altitude work safety monitoring.

[0003] Current methods for supervising safety in high-altitude operations primarily rely on manual patrols, fixed camera monitoring, and on-site observation by safety officers. While manual patrols can be based on experience, patrol personnel cannot continuously monitor the entire work area, and blind spots and delays in monitoring can easily occur when multiple floors, platforms, and personnel are working simultaneously. Although fixed camera monitoring can continuously record the work area, traditional video surveillance typically only provides video recordings, requiring managers to manually review the footage and determine if any violations are occurring. It cannot automatically identify changes in workers' postures or edge-risk conditions. When violations are short-lived, occur in concealed locations, or are obstructed by equipment, relying solely on manual video review can easily lead to missed detections and make it difficult to promptly identify dangerous actions by workers at height. Summary of the Invention

[0004] One objective of this invention is to propose a method for identifying violations by high-altitude workers based on posture capture. This invention fully utilizes computer vision, human posture capture, scene baseline reconstruction, posture-boundary coupled projection, and an improved CleverCatch model, and has the advantages of high recognition accuracy, low false alarm rate, strong timeliness of supervision, and applicability to complex high-altitude work scenarios.

[0005] A method for identifying violations by high-altitude workers based on attitude capture according to an embodiment of the present invention includes the following steps: Collect on-site image data and work space calibration data of the high-altitude work area, and preprocess the on-site image data to generate a high-altitude work image frame sequence; Based on the workspace calibration data, the high-altitude work area is reconstructed to generate high-altitude work scene benchmark data. Based on the high-altitude operation image frame sequence, human body contour segmentation and personnel region anchoring are performed on the workers to generate a personnel posture observation sequence. Human posture is captured from the personnel posture observation sequence to generate a sequence of worker posture key points; The posture trajectory of the operator's posture key point sequence is refined to generate the operator's posture trajectory data; The attitude-boundary coupled projection is performed between the personnel attitude trajectory data and the high-altitude operation scenario baseline data to obtain attitude-boundary coupled data. The attitude boundary coupling data is input into the improved CleverCatch model to generate candidate violation data. The candidate violation data are subjected to time-series continuity verification and attitude stability verification to generate valid violation data and obtain the violation identification results of high-altitude workers.

[0006] Optionally, image acquisition equipment is deployed in the high-altitude work area to continuously capture images of the work platform, guardrails, edge areas, personnel activity areas, and high-altitude workers, thereby obtaining image data of the work site and acquiring work space calibration data corresponding to the high-altitude work area. The preprocessing includes frame sequence extraction, invalid frame removal, image denoising, distortion correction, illumination equalization, and timestamp synchronization processing.

[0007] Optionally, the generation of the high-altitude operation scenario reference data specifically includes: The coordinate system of the workspace calibration data is processed to generate spatial boundary calibration data. Image coordinate transformation is performed on the spatial boundary calibration data to generate image boundary calibration data corresponding to the high-altitude operation image frame sequence; The image boundary calibration data is parsed for boundary type and reconstructed for boundary morphology to generate a set of scene reference units; Based on the positional relationships between the scene reference units in the scene reference unit set, a scene reference structure is constructed; By associating the scene baseline structure with the image boundary calibration data, high-altitude operation scene baseline data is generated.

[0008] Optionally, the generation of the personnel posture observation sequence specifically includes: Configure frame numbers for each work site image frame in the high-altitude operation image frame sequence to generate frame number data; Based on frame number data, personnel foreground is extracted from each work site image frame in the high-altitude operation image frame sequence to generate personnel pixel regions. Edge convergence and hole filling processes are applied to the pixel regions of people to generate human body contour segmentation data. The contour boundaries of the human body contour segmentation data are extracted to generate personnel region anchor boxes, and the personnel region anchor boxes are associated with frame number data to generate worker region anchor data. The anchor boxes of personnel regions in adjacent work site image frames are continuously matched, personnel identity identifiers are assigned to the continuously matched personnel region anchor boxes, and the personnel region anchor data are time-series collected to generate a personnel posture observation sequence.

[0009] Optionally, the generation of the operator posture key point sequence specifically includes: Based on the personnel posture observation sequence, personnel posture image regions are extracted to generate a personnel posture image region sequence; The scale normalization process is performed on the personnel pose image region sequence, and the correspondence between the region coordinates before and after the scale normalization process is recorded to generate pose capture input data and region coordinate reconstruction data. The posture capture input data is input into the human posture capture network for thermal response extraction, generating joint thermal response data. Peak positions are extracted and confidence levels are filtered for joint thermal response data. The normalized coordinates of each human joint in the posture capture input data are determined. Combined with the regional coordinate restoration data, the normalized coordinates of the joints are restored to the image coordinate system of the corresponding work site image frame to generate single-frame posture joint data. The posture key point data of the same worker in continuous work site image frames are arranged in time sequence to generate the worker posture key point sequence.

[0010] Optionally, the generation of the personnel posture trajectory data specifically includes: Arrange the single-frame attitude joint data in the operator's attitude joint sequence in a temporal order to generate joint timing data; The image coordinates of the same human joint in the time series data of joint points are concatenated across frames in continuous work site image frames to generate joint point motion trajectory data. Based on the motion trajectory data of the joints, the changes in torso orientation, limb extension, foot support, and center of gravity of the same worker in continuous work site image frames are extracted to generate posture state feature data. The posture trajectory is refined from the posture state feature data to generate personnel posture trajectory data.

[0011] Optionally, obtaining the attitude boundary coupling data specifically includes: The posture trajectory data of the same worker in the same work site image frame is matched with the reference data of the high-altitude work scene to generate posture scene matching data. Generate posture projection element data based on posture scene matching data; Generate boundary projection reference data based on posture scene matching data; The attitude projection element data is projected onto the image coordinate system corresponding to the boundary projection reference data to generate attitude boundary position data and form attitude boundary coupling data.

[0012] Optionally, the generation of the candidate violation data specifically includes: The improved CleverCatch model incorporates attitude boundary coupling data as input. This improved model includes an attitude boundary coupling shaping module, a near-edge attitude risk focusing module, a violation attitude link tracking module, and a candidate violation behavior extraction module. The improvements are as follows: the traditional CleverCatch model captures and classifies changes in the target state in the input data, focusing on identifying changes in target behavior in continuous data. The improved CleverCatch model uses attitude boundary coupling data as input, performing coupling shaping on the positional relationships between the torso orientation trajectory, limb extension trajectory, foot support trajectory, human center of gravity trajectory, and the scene's baseline structure. It also introduces an attitude risk vortex traction mechanism, generating near-edge attitude risk features based on the direction, magnitude, and duration of changes in orientation offset, extension overshoot, support deviation, and center of gravity approach. Furthermore, it performs cross-frame risk vortex residual tracking to generate violation attitude link data, forming candidate violation behavior data. In the posture boundary coupling shaping module, the positional relationship between the torso orientation trajectory, limb extension trajectory, foot support trajectory, human body center trajectory and scene reference structure in the posture boundary coupling data is coupled and shaped to generate posture boundary coupling features. In the edge posture risk focusing module, a posture risk eddy traction mechanism is introduced. Based on the posture boundary coupling characteristics, posture risk eddy response data is constructed. Based on the posture risk eddy response data, risk focusing processing is performed on posture changes that continuously converge toward the edge reference area, limb changes that continuously cross the guardrail reference line, foot support changes that continuously deviate from the personnel's movable reference area, and changes in the human body's center of gravity that continuously approach the edge reference area, thereby generating edge posture risk characteristics. In the violation posture link tracking module, cross-frame risk eddy residual tracking processing is performed on the edge posture risk characteristics. According to the frame number data, the continuous edge posture risk characteristics corresponding to the same personnel identification are read to generate violation posture link nodes. Based on the start frame, continuation frame and fallback frame of the violation posture link nodes in the continuous work site image frames, the edge posture risk change process of the same worker is linked together to generate violation posture link data. In the candidate violation extraction module, candidate violation types are extracted based on the violation posture link data, and candidate violation types, personnel identification, and frame number data are associated to generate candidate violation data.

[0013] Optionally, obtaining the results of the identification of violations by high-altitude workers specifically includes: Perform time-series continuous verification on candidate violation data to determine valid time-series candidate violation data; Based on the time-series valid candidate violation data, the frame number data of the same personnel identification and the same candidate violation type in the continuous operation site image frames are read, and the attitude stability is verified according to the maintenance state of the candidate violation type and the continuous state of the frame number data in the continuous operation site image frames to generate stable candidate violation data. The time-series valid candidate violation data that maintains the same candidate violation type and has continuous frame number data in the stable candidate violation data are identified as valid violation data, thus obtaining the violation identification results of high-altitude workers.

[0014] The beneficial effects of this invention are: This invention collects on-site image data and work space calibration data of high-altitude work areas, preprocesses the on-site image data to generate a sequence of high-altitude work image frames, and simultaneously reconstructs the scene benchmark of the high-altitude work area based on the work space calibration data to generate high-altitude work scene benchmark data. This allows the subsequent recognition process to move beyond relying solely on the physical features of personnel in a single video frame, and instead analyze spatial benchmarks such as the work platform, guardrails, edge areas, and areas where personnel can move within the high-altitude work area. This improves the scene adaptability for recognizing personnel violations in high-altitude work scenarios and overcomes the limitations of ordinary video surveillance, which can only record images and struggles to understand the boundary relationships of the work space.

[0015] This invention generates a sequence of worker posture observations by segmenting the human body contours and anchoring the personnel region in a sequence of high-altitude operation image frames. Further, it performs human posture capture to generate a sequence of worker posture key points. This allows for relatively stable extraction of posture data from the same worker even under conditions of multiple workers working simultaneously, varying worker scales, partial occlusion, and complex background components. Compared to methods relying solely on target detection boxes or manual review of monitoring footage, this invention can more accurately obtain changes in worker body posture, providing a continuous and reliable data foundation for subsequent violation identification.

[0016] This invention generates personnel posture trajectory data by refining the posture joint sequence of workers, and then performs posture-boundary coupling projection on the personnel posture trajectory data and high-altitude operation scene reference data to obtain posture boundary coupling data. This allows for correlation analysis of the positional relationships between the torso orientation trajectory, limb extension trajectory, foot support trajectory, and human center of gravity trajectory and the scene reference structure. This effectively distinguishes normal work actions such as bending over, turning, and reaching out for work from dangerous behaviors such as leaning out of bounds, illegally crossing guardrails, abnormal standing near edges, abnormal center of gravity shift, and unbalanced tilting, reducing misidentification caused by similar single-frame images.

[0017] This invention improves the CleverCatch model by inputting attitude boundary coupling data into the model and generating candidate violation data through attitude boundary coupling shaping, edge-prone attitude risk focusing, violation attitude link tracking, and candidate violation behavior extraction processing. This enables continuous analysis of the attitude change process of high-altitude workers. In particular, through the attitude risk vortex traction mechanism and cross-frame risk vortex residual tracking processing, it can highlight risk changes such as continuous convergence towards the edge area, continuous crossing of the guardrail baseline, continuous deviation from the personnel's movable area, and continuous approach of the human body's center of gravity towards the edge area. This allows violation behavior identification to go beyond single-frame anomaly judgment and instead combine risk change trends in consecutive frames for identification, thereby improving the accuracy and stability of identification.

[0018] This invention further performs temporal continuity verification and posture stability verification on candidate violation data to generate valid violation data and obtain violation identification results for high-altitude workers. It can eliminate invalid candidate violations caused by short-term occlusion, joint vibration, normal posture switching, and momentary misidentification. This reduces false alarm and false negative rates, improves the reliability of high-altitude violation identification results, and enables on-site supervisors to more timely and accurately grasp the type of violation, personnel identification, and duration of violation for high-altitude workers. This provides automated, refined, and traceable safety supervision methods for high-altitude work scenarios such as construction sites, power maintenance, bridge maintenance, and curtain wall installation. Attached Figure Description

[0019] 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: Fig. 1 This is an overall flowchart of a method for identifying violations by high-altitude workers based on attitude capture, as proposed in this invention. Fig. 2 This is a schematic diagram illustrating the construction of attitude boundary coupling data for a method for identifying violations by high-altitude workers based on attitude capture proposed in this invention. Fig. 3 This is a schematic diagram of the improved CleverCatch model, which is a method for identifying violations by high-altitude workers based on attitude capture, as proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figs. 1-3 A method for identifying violations by high-altitude workers based on attitude capture includes the following steps: Collect on-site image data and work space calibration data of the high-altitude work area, and preprocess the on-site image data to generate a high-altitude work image frame sequence; Based on the workspace calibration data, the high-altitude work area is reconstructed to generate high-altitude work scene benchmark data. Based on the high-altitude operation image frame sequence, human body contour segmentation and personnel region anchoring are performed on the workers to generate a personnel posture observation sequence. Human posture is captured from the personnel posture observation sequence to generate a sequence of worker posture key points; The posture trajectory of the operator's posture key point sequence is refined to generate the operator's posture trajectory data; The attitude-boundary coupled projection is performed between the personnel attitude trajectory data and the high-altitude operation scenario baseline data to obtain attitude-boundary coupled data. The attitude boundary coupling data is input into the improved CleverCatch model to generate candidate violation data. The candidate violation data are subjected to time-series continuity verification and attitude stability verification to generate valid violation data and obtain the violation identification results of high-altitude workers.

[0022] In this embodiment, image acquisition equipment is deployed in the high-altitude work area to continuously capture images of the work platform, guardrails, edge areas, personnel activity areas, and high-altitude workers, thereby obtaining image data of the work site and acquiring work space calibration data corresponding to the high-altitude work area. The work space calibration data includes the installation height of the image acquisition equipment, shooting angle, lens intrinsic parameters, lens extrinsic parameters, work platform boundary position, guardrail position, edge area position, and personnel activity area position. Preprocessing includes frame sequence extraction, invalid frame removal, image denoising, distortion correction, illumination equalization, and timestamp synchronization processing.

[0023] In this embodiment, the generation of reference data for high-altitude operation scenarios specifically includes: The coordinate system of the workspace calibration data is processed to generate spatial boundary calibration data. The specific process for generating spatial boundary calibration data is as follows: Read the image acquisition device's intrinsic and extrinsic parameters, work platform boundary data, guardrail boundary data, adjacent area boundary data, and personnel-accessible area boundary data from the work space calibration data; establish a high-altitude work area coordinate system using the work platform plane as the reference plane; based on the image acquisition device's intrinsic and extrinsic parameters, uniformly convert the work platform boundary data, guardrail boundary data, adjacent area boundary data, and personnel-accessible area boundary data to the high-altitude work area coordinate system; perform coordinate sorting, boundary point deduplication, and boundary point continuity processing on the converted work platform boundary data, guardrail boundary data, adjacent area boundary data, and personnel-accessible area boundary data to generate spatial boundary calibration data. Image coordinate transformation is performed on the spatial boundary calibration data to generate image boundary calibration data corresponding to the high-altitude operation image frame sequence; The image coordinate transformation process is as follows: Read the intrinsic and extrinsic parameter data of the image acquisition device from the work space calibration data. The extrinsic parameter data includes the rotation and translation parameters of the image acquisition device relative to the coordinate system of the high-altitude work area. Based on the rotation parameters, perform orientation transformation on the coordinates of each boundary point in the spatial boundary calibration data. Based on the translation parameters, perform position translation on the coordinates of each boundary point after orientation transformation, transforming each boundary point from the high-altitude work area coordinate system to the image acquisition device coordinate system. Based on the focal length data, principal point coordinate data, and pixel scale data in the image acquisition device's intrinsic parameter data, project each boundary point in the image acquisition device coordinate system to the image coordinate system corresponding to the high-altitude work image frame sequence, obtaining the pixel position of each boundary point in the image coordinate system. According to the type of work platform boundary, guardrail boundary, adjacent area boundary, and personnel-accessible area boundary, classify, collect, and arrange the projected boundary points sequentially to generate image boundary calibration data corresponding to the high-altitude work image frame sequence. The image boundary calibration data is parsed for boundary type and reconstructed for boundary morphology to generate a set of scene reference units; The specific process for generating the scene reference unit set is as follows: Read the image boundary calibration data, and according to the boundary source identifier of each boundary point in the image boundary calibration data, parse the boundary type of each boundary point to obtain the work platform boundary point group, guardrail boundary point group, adjacent area boundary point group, and personnel-accessible area boundary point group; perform boundary point sequence arrangement and area closure processing on the work platform boundary point group to generate the work platform reference area; perform boundary point sequence arrangement and line segment connectivity processing on the guardrail boundary point group to generate the guardrail baseline; perform boundary point sequence arrangement and area closure processing on the adjacent area boundary point group to generate the adjacent reference area; perform boundary point sequence arrangement and area closure processing on the personnel-accessible area boundary point group to generate the personnel-accessible reference area; use the work platform reference area, guardrail baseline, adjacent area reference area, and personnel-accessible reference area as scene reference units, and store them in association according to their corresponding boundary source identifiers to generate the scene reference unit set. Based on the positional relationships between the scene reference units in the scene reference unit set, a scene reference structure is constructed; The construction process of the scene reference structure is as follows: Read each scene reference unit in the scene reference unit set; based on the positional distribution of the work platform reference area, guardrail reference line, edge reference area, and personnel mobility reference area in the same image coordinate system, analyze the inclusion, adjacency, spacing, and overlap relationships between each scene reference unit; use the work platform reference area as the load-bearing reference for the high-altitude work area, the guardrail reference line as the protection reference for the high-altitude work area, the edge reference area as the danger boundary reference for the high-altitude work area, and the personnel mobility reference area as the safety activity reference for high-altitude workers; determine the protection position relationship based on the adjacency relationship between the guardrail reference line and the edge reference area, and determine the safety activity position relationship based on the spacing relationship between the personnel mobility reference area and the edge reference area; structurally associate the scene reference unit set, protection position relationship, and safety activity position relationship to form the scene reference structure. By associating the scene reference structure with the image boundary calibration data, high-altitude operation scene reference data is generated; The specific process for generating reference data for high-altitude operations scenarios is as follows: First, read the scenario reference structure and image boundary calibration data. Then, use the image boundary calibration data as the image position reference for the scenario reference structure within the high-altitude operations image frame sequence. Next, associate the scenario reference unit set, protection position relationship, and safe activity position relationship in the scenario reference structure with the corresponding boundary point positions in the image boundary calibration data, generating scenario reference associated data that includes scenario reference units, boundary point positions, protection position relationships, and safe activity position relationships. Finally, index the scenario reference associated data according to the frame number of the high-altitude operations image frame sequence, enabling the operation site image frames under the same frame number to call the corresponding scenario reference associated data. The indexed scenario reference associated data is then used as the high-altitude operations scenario reference data.

[0024] In this embodiment, the generation of the personnel attitude observation sequence specifically includes: Configure frame numbers for each work site image frame in the high-altitude operation image frame sequence to generate frame number data; The process of generating frame number data is as follows: read each work site image frame in the high-altitude operation image frame sequence and read the image acquisition time corresponding to each work site image frame; sort each work site image frame according to the order of image acquisition time; assign a unique frame number to each work site image frame according to the sorting result, and bind the frame number with the corresponding work site image frame's image acquisition time; index and organize the bound frame numbers and image acquisition times to generate frame number data. Based on frame number data, personnel foreground is extracted from each work site image frame in the high-altitude operation image frame sequence to generate personnel pixel regions. The process of generating personnel pixel regions is as follows: Read each work site image frame from the high-altitude operation image frame sequence and determine the processing order of each work site image frame based on the frame number data; perform background suppression processing on each work site image frame to weaken the background pixel response corresponding to the work platform, guardrail, equipment, and fixed components; extract the personnel foreground from the background-suppressed work site image frames and identify the foreground pixels belonging to the workers; group the foreground pixels belonging to the same worker into connected regions, remove discrete noise regions with fewer than a preset pixel count threshold, and retain the foreground connected regions that satisfy the continuity of the personnel's shape; use the retained foreground connected regions as the personnel pixel regions. Edge convergence and hole filling processes are applied to the pixel regions of people to generate human body contour segmentation data. The specific process for generating human contour segmentation data is as follows: Edge convergence processing is performed on the outer pixel boundaries of the human pixel region, merging discrete foreground pixels at the boundary into adjacent continuous foreground pixel regions, removing spiky and isolated pixels from the edges of the human pixel region; hole filling processing is performed on the edge-converged human pixel region, filling the empty pixel regions within the human pixel region that were not identified as foreground pixels and are surrounded by continuous foreground pixels; contour smoothing processing is performed on the human pixel region after hole filling processing to form a human contour region with continuous outer boundaries and complete internal regions; the human contour region is then associated with the corresponding frame number data to generate human contour segmentation data. The contour boundaries of the human body contour segmentation data are extracted to generate personnel region anchor boxes, and the personnel region anchor boxes are associated with frame number data to generate worker region anchor data. The specific process for generating worker area anchoring data is as follows: First, the human contour region in each work site image frame is determined based on the human contour segmentation data. Second, the outer boundary of the human contour region is extracted to obtain the leftmost, rightmost, topmost, and bottommost boundary points of the human contour region in the image coordinate system. Third, based on the leftmost, rightmost, topmost, and bottommost boundary points, a rectangular region enclosing the human contour region is determined, and this rectangular region is used as the worker area anchoring frame. Fourth, the worker area anchoring frame is associated with the frame number data of the corresponding work site image frame, and the position and size data of the worker area anchoring frame in the image coordinate system are recorded to generate worker area anchoring data. The anchor boxes of personnel regions in adjacent work site image frames are continuously matched, personnel identity identifiers are assigned to the continuously matched personnel region anchor boxes, and the personnel region anchor data are time-series collected to generate personnel posture observation sequences. The generation process of the personnel posture observation sequence is as follows: The temporal sequence between adjacent work site image frames is determined according to frame number data. Continuity matching is performed on the personnel region anchor boxes in adjacent work site image frames. Continuity matching refers to determining whether two personnel region anchor boxes belong to the same worker based on changes in the center position, size, and overlap of the personnel region anchor boxes in adjacent work site image frames. When the personnel region anchor boxes in adjacent work site image frames satisfy continuity matching, the personnel region anchor boxes are divided into continuous anchor boxes corresponding to the same worker, and the same personnel identification is assigned to the continuous anchor boxes. The worker region anchor data is classified according to the personnel identification, and the worker region anchor data corresponding to the same personnel identification is arranged chronologically according to frame number data, generating a personnel posture observation sequence representing the position and region changes of the same worker in consecutive work site image frames.

[0025] In this embodiment, the generation of the operator's posture key point sequence specifically includes: Based on the personnel posture observation sequence, personnel posture image regions are extracted to generate a personnel posture image region sequence; The process of generating the personnel posture image region sequence is as follows: Based on the personnel posture observation sequence, the temporal order of the workers in the continuous work site image frames is determined according to the frame number data; the personnel region anchoring boxes in the personnel region anchoring data are read, and the work site image frames are cropped according to the position range of the personnel region anchoring boxes in the corresponding work site image frames to obtain personnel posture image regions containing the outline of the workers' bodies; the personnel posture image regions corresponding to the same personnel identification are arranged sequentially according to the frame number data to generate the personnel posture image region sequence. The scale normalization process is performed on the personnel pose image region sequence, and the correspondence between the region coordinates before and after the scale normalization process is recorded to generate pose capture input data and region coordinate reconstruction data. The process of generating pose capture input data and region coordinate reconstruction data is as follows: Each person pose image region in the personnel pose image region sequence undergoes scale normalization processing to convert personnel pose image regions of different sizes into pose capture input images of uniform size; during the scale normalization process, the original region width, original region height, and original upper-left corner coordinates of each personnel pose image region before scale normalization are recorded, as well as the input region width and input region height after scale normalization; based on the proportional relationship between the original region width and the input region width, the proportional relationship between the original region height and the input region height, and the original upper-left corner coordinates, region coordinate reconstruction data is generated; the pose capture input images after scale normalization are associated and arranged according to personnel identification and frame number data to generate pose capture input data. The posture capture input data is input into the human posture capture network for thermal response extraction, generating joint thermal response data. The specific process for generating joint thermal response data is as follows: The posture capture input data is input into a human posture capture network, which includes a posture feature extraction layer, a multi-scale feature fusion layer, a joint response decoding layer, and a thermal response output layer. The posture feature extraction layer extracts features from the human body shape, limb edges, and local textures in the posture capture input data to generate basic posture features. The multi-scale feature fusion layer performs scale fusion on the local features corresponding to the head, shoulders, elbows, wrists, hips, knees, and ankles in the basic posture features to generate posture fusion features. The joint response decoding layer analyzes the joint position response of the posture fusion features to obtain the position response distribution of each human joint in the posture capture input data. The thermal response output layer converts the position response distribution of each human joint into a corresponding joint heatmap, and associates each joint heatmap according to personnel identification and frame number data to generate joint thermal response data. Peak positions are extracted and confidence levels are filtered for joint thermal response data. The normalized coordinates of each human joint in the posture capture input data are determined. Combined with the regional coordinate restoration data, the normalized coordinates of the joints are restored to the image coordinate system of the corresponding work site image frame to generate single-frame posture joint data. The specific process for generating single-frame attitude joint data is as follows: Peak positions are extracted from the joint heatmap corresponding to each human joint in the joint thermal response data. The pixel position with the largest response value in the joint heatmap is determined as the initial normalized coordinate of the human joint in the attitude capture input data, and the maximum response value is used as the joint confidence level of the corresponding human joint. The joint confidence level is compared with a preset joint confidence threshold, and the normalized coordinates of joints whose confidence levels reach the preset threshold are retained. Based on the original region width, original region height, original region top-left corner coordinates, input region width, and input region height in the region coordinate reconstruction data, the retained normalized coordinates of the joints are restored from the normalized coordinate system corresponding to the attitude capture input data to the image coordinate system of the corresponding work site image frame. The restored image coordinates, joint confidence levels, personnel identification, and frame number data of each human joint are associated to generate single-frame attitude joint data. The posture key point data of the same worker in continuous work site image frames are arranged in time sequence to generate the worker posture key point sequence.

[0026] In this embodiment, the generation of personnel posture trajectory data specifically includes: Arrange the single-frame attitude joint data in the operator's attitude joint sequence in a temporal order to generate joint timing data; The image coordinates of the same human joint in the time series data of joint points are concatenated across frames in continuous work site image frames to generate joint point motion trajectory data. The specific process for generating joint motion trajectory data is as follows: First, determine the timing data of joints corresponding to the same worker based on their identification, and then determine the temporal order of the same worker in continuous work site image frames based on frame number data. For the same human joint, sequentially read the image coordinates of the human joint in continuous work site image frames, and concatenate the image coordinates of the human joint across frames according to the order of frame number data to generate a single joint motion trajectory. Finally, associate the single joint motion trajectories corresponding to the same worker identification to generate joint motion trajectory data. Based on the motion trajectory data of the joints, the changes in torso orientation, limb extension, foot support, and center of gravity of the same worker in continuous work site image frames are extracted to generate posture state feature data. The specific process for generating posture state feature data is as follows: Based on the relative position changes between human joints located in the torso area in the joint motion trajectory data, determine the torso orientation change of the same worker in continuous work site image frames; based on the distance and direction changes between human joints located in the limb areas in the joint motion trajectory data, determine the limb extension changes of the same worker in continuous work site image frames; based on the position changes of human joints located in the feet area in the joint motion trajectory data in continuous work site image frames, determine the foot support changes of the same worker in continuous work site image frames; based on the image coordinate distribution of each human joint in the joint motion trajectory data, calculate the change of the center of gravity of the same worker in continuous work site image frames; associate the torso orientation change, limb extension change, foot support change, and center of gravity change according to the personnel identification and frame number data to generate posture state feature data; The posture trajectory is refined from the posture state feature data to generate personnel posture trajectory data; The process of posture trajectory condensation is as follows: First, determine the posture state feature data corresponding to the same worker according to their identification, and determine the temporal order of the posture state feature data according to the frame number data. Second, perform continuous frame aggregation on the changes in torso orientation, limb extension, foot support, and center of gravity of the same worker in continuous work site image frames, removing short-term jump data caused by single-frame joint offsets, and retaining the posture change data that continuously appears in the continuous work site image frames. Third, extract the torso orientation trajectory, limb extension trajectory, foot support trajectory, and center of gravity trajectory of the same worker based on the retained posture change data. Fourth, associate the torso orientation trajectory, limb extension trajectory, foot support trajectory, and center of gravity trajectory according to the worker's identification and frame number data to generate personnel posture trajectory data.

[0027] In this embodiment, obtaining the attitude boundary coupling data specifically includes: The posture trajectory data of the same worker in the same work site image frame is matched with the reference data of the high-altitude work scene to generate posture scene matching data. The specific process for generating posture scene matching data is as follows: Based on the personnel identification, determine the personnel posture trajectory data corresponding to the same worker, and based on the frame number data, determine the posture trajectory position of the worker in the corresponding work site image frame; read the high-altitude work scene reference data corresponding to the same frame number data, and perform intra-frame correspondence matching between the personnel posture trajectory data in the work site image frame and the corresponding image boundary calibration data and scene reference structure; associate the personnel identification, frame number data, personnel posture trajectory data, image boundary calibration data, and scene reference structure after completing the intra-frame correspondence matching to generate posture scene matching data. Generate posture projection element data based on posture scene matching data; The specific process of generating posture projection element data is as follows: extract the torso orientation trajectory, limb extension trajectory, foot support trajectory, and body center trajectory of the same worker from the personnel posture trajectory data in the posture scene matching data in the corresponding work site image frame; determine the work site image frame corresponding to the torso orientation trajectory, limb extension trajectory, foot support trajectory, and body center trajectory according to the frame number data; read the image coordinates of the torso orientation trajectory, limb extension trajectory, foot support trajectory, and body center trajectory in the work site image frame, and associate the image coordinates with the personnel identification and frame number data to generate posture projection element data; Generate boundary projection reference data based on posture scene matching data; The specific process of generating boundary projection reference data is as follows: Based on the attitude scene matching data, the corresponding work site image frame is determined according to the frame number data, and the image boundary calibration data corresponding to the work site image frame is read; the scene reference unit set, protection position relationship and safety activity position relationship are extracted from the scene reference structure; the image coordinate position of the scene reference unit set in the corresponding work site image frame is determined according to the image boundary calibration data, and the scene reference unit set, protection position relationship, safety activity position relationship and image coordinate position are associated according to the frame number data to generate boundary projection reference data; The attitude projection element data is projected onto the image coordinate system corresponding to the boundary projection reference data to generate attitude boundary position data and form attitude boundary coupling data. The process of forming posture boundary coupling data is as follows: align the posture projection element data of the same worker in the same work site image frame with the boundary projection reference data; based on the image coordinate system corresponding to the boundary projection reference data, compare the image coordinates corresponding to the torso orientation trajectory, limb extension trajectory, foot support trajectory, and human center of gravity trajectory with the image coordinate positions corresponding to the scene reference unit set to generate posture boundary position data; based on the posture boundary position data, associate the positional relationships between the torso orientation trajectory and the scene reference structure, the limb extension trajectory and the scene reference structure, the foot support trajectory and the scene reference structure, and the human center of gravity trajectory and the scene reference structure to form posture boundary coupling data.

[0028] In this embodiment, the generation of candidate violation data specifically includes: The improved CleverCatch model incorporates attitude boundary coupling data as input. This improved model includes an attitude boundary coupling shaping module, a near-edge attitude risk focusing module, a violation attitude link tracking module, and a candidate violation behavior extraction module. The improvements are as follows: the traditional CleverCatch model captures and classifies changes in the target state in the input data, focusing on identifying changes in target behavior in continuous data. The improved CleverCatch model uses attitude boundary coupling data as input, performing coupling shaping on the positional relationships between the torso orientation trajectory, limb extension trajectory, foot support trajectory, human center of gravity trajectory, and the scene's baseline structure. It also introduces an attitude risk vortex traction mechanism, generating near-edge attitude risk features based on the direction, magnitude, and duration of changes in orientation offset, extension overshoot, support deviation, and center of gravity approach. Furthermore, it performs cross-frame risk vortex residual tracking to generate violation attitude link data, forming candidate violation behavior data. The improved CleverCatch model consists of a sequentially connected attitude boundary coupling shaping module, an edge attitude risk focusing module, a violation attitude link tracking module, and a candidate violation behavior extraction module. Attitude boundary coupling data serves as the model input, formatted as multi-frame attitude boundary data arranged according to personnel identification and frame number. The attitude boundary coupling shaping module converts the attitude boundary coupling data into attitude boundary coupling features, whose dimensions include orientation offset, extension overrun, support deviation, and center of gravity approach. The edge attitude risk focusing module receives the attitude boundary coupling features and generates edge attitude risk features based on the attitude risk vortex traction mechanism. These features include orientation offset, extension overrun, support deviation, center of gravity approach, and edge risk focus value. The violation attitude link tracking module receives the edge attitude risk features and performs cross-frame risk vortex residual tracking according to frame number data to generate violation attitude link data. The candidate violation behavior extraction module receives the violation attitude link data and generates candidate violation behavior data based on the continuous changes in the violation attitude link nodes. The training data for the improved CleverCatch model comes from on-site image data collected in high-altitude work areas, work space calibration data, and sample data of violations by high-altitude workers formed by manual review. The annotation method is to annotate the candidate violation type, start frame, continuation frame, and fall frame of the posture boundary coupled data of the same worker according to the worker's identity and frame number data, and simultaneously annotate the risk change status corresponding to the orientation offset, extension over-boundary amount, support deviation amount, and center of gravity approach amount. The loss function includes candidate violation type classification loss, violation posture link node localization loss, and edge risk focus value regression loss, which are used to constrain the model to learn the violation type, continuous frame link position, and edge risk change degree at the same time. The training parameters were set as follows: 100 training epochs, 32 samples per batch, initial learning rate of 0.001, learning rate decay coefficient of 0.1, learning rate decay interval of 20 epochs, weight decay coefficient of 0.0005, momentum coefficient of 0.9, training sample to validation sample ratio of 8:2, and continuous frame input length of 16 frames. During training, the change in validation set loss was used as the basis for model updates. When the decrease in validation set loss was less than 0.001 in 10 consecutive training epochs, and the accuracy of candidate violation behavior type recognition reached more than 95%, the improved CleverCatch model was considered to have reached convergence. In the posture boundary coupling shaping module, the positional relationship between the torso orientation trajectory, limb extension trajectory, foot support trajectory, human body center trajectory and scene reference structure in the posture boundary coupling data is coupled and shaped to generate posture boundary coupling features. The specific process of generating posture boundary coupling features is as follows: Based on personnel identification and frame number data, the torso orientation trajectory, limb extension trajectory, foot support trajectory, and center of gravity trajectory of the same worker in the same work site image frame are synchronously read; the orientation offset of the torso orientation trajectory relative to the scene reference structure, the extension overshoot of the limb extension trajectory relative to the scene reference structure, the support deviation of the foot support trajectory relative to the scene reference structure, and the center of gravity approach of the center of gravity trajectory relative to the scene reference structure are calculated respectively; the orientation offset, extension overshoot, support deviation, and center of gravity approach are aligned and arranged according to the same personnel identification and the same frame number data to form a posture boundary coupling vector; based on the numerical changes of the posture boundary coupling vector in consecutive frames, the orientation offset, extension overshoot, support deviation, and center of gravity approach are correlated in time sequence to generate posture boundary coupling features that characterize the coupling change state between the posture trajectory of the same worker and the scene reference structure. In the edge posture risk focusing module, a posture risk eddy traction mechanism is introduced. Based on the posture boundary coupling characteristics, posture risk eddy response data is constructed. Based on the posture risk eddy response data, risk focusing processing is performed on posture changes that continuously converge toward the edge reference area, limb changes that continuously cross the guardrail reference line, foot support changes that continuously deviate from the personnel's movable reference area, and changes in the human body's center of gravity that continuously approach the edge reference area, thereby generating edge posture risk characteristics. The specific process for generating edge posture risk features is as follows: Based on posture boundary coupling features, an posture risk vortex traction mechanism is introduced. The direction of change, amplitude of change, and number of consecutive frames are extracted for the orientation offset, extension over-boundary amount, support deviation, and center of gravity approach of the same worker in continuous work site image frames. The changes in orientation offset continuously increasing towards the edge reference area, extension over-boundary amount continuously increasing relative to the guardrail reference line, support deviation continuously increasing relative to the worker's movable reference area, and center of gravity approach relatively increasing relative to the edge reference area are used as the objects of posture risk vortex traction. Based on the changing direction, amplitude, and duration of the attitude risk vortex-driven object, attitude risk vortex response data is generated. Risk focusing processing is then performed on this data. This involves assigning risk focusing weights to orientation offset, extension overrun, support deviation, and center of gravity approach based on the attitude risk vortex response data. These weights are then weighted and fused with the corresponding orientation offset, extension overrun, support deviation, and center of gravity approach values ​​to obtain the edge risk focusing value. This edge risk focusing value is then associated with personnel identification and frame number data to generate edge attitude risk characteristics. In the violation posture link tracking module, cross-frame risk eddy residual tracking processing is performed on the edge posture risk characteristics. According to the frame number data, the continuous edge posture risk characteristics corresponding to the same personnel identification are read to generate violation posture link nodes. Based on the start frame, continuation frame and fallback frame of the violation posture link nodes in the continuous work site image frames, the edge posture risk change process of the same worker is linked together to generate violation posture link data. The specific process for generating violation posture link data is as follows: First, determine the continuous edge-prone posture risk characteristics corresponding to the same worker based on their identification, and determine the temporal order of these characteristics based on frame number data. Second, perform cross-frame risk vortex residual tracking processing on the continuous edge-prone posture risk characteristics. This processing involves calculating the difference between the edge-prone risk focus value in the current work site image frame and the edge-prone risk focus value in the previous work site image frame to determine the cross-frame residual change of the edge-prone risk focus value. Combined with the continuous changes in orientation offset, extension over-limit, support deviation, and center of gravity approach, determine the risk residual state. Third, when the risk residual state changes from a non-persistent state to a persistent state, the corresponding work site image frame is designated as the starting frame. When the risk residual state remains persistent, the corresponding work site image frame is designated as the continuing frame. When the risk residual state changes from a persistent state to a non-persistent state, the corresponding work site image frame is designated as the fallback frame. Fourth, use the starting frame, continuing frame, and fallback frame as violation posture link nodes, and connect these nodes according to frame number data to generate violation posture link data. In the candidate violation extraction module, candidate violation types are extracted based on the violation posture link data, and candidate violation types, personnel identification and frame number data are associated to generate candidate violation data; The specific process for generating candidate violation data is as follows: First, determine the violation posture link data corresponding to the same worker based on their identification, and then determine the time sequence of the violation posture link nodes based on the frame number data. Second, based on the edge risk focus value, orientation offset, extension overstepping, support deviation, and center of gravity approach value corresponding to the violation posture link nodes, extract the behavior type from the posture risk changes of the same worker between the initial frame, continuation frame, and fallback frame. Behavior type extraction refers to determining the corresponding candidate violation behavior type based on the dominant change relationship of different risk quantities in the violation posture link data. When the orientation offset dominates the change in consecutive violation posture link nodes, the candidate violation behavior type is determined to be "body overstepping." When the outbound behavior and the amount of extension exceeding the boundary dominate the changes in the continuous violation posture link nodes, the candidate violation behavior type is determined to be illegal crossing of the guardrail. When the amount of support deviation dominates the changes in the continuous violation posture link nodes, the candidate violation behavior type is determined to be abnormal standing near the edge. When the amount of center of gravity approach dominates the changes in the continuous violation posture link nodes, the candidate violation behavior type is determined to be abnormal center of gravity shift. When the near-edge risk focus value continuously increases in the continuous violation posture link nodes and the orientation shift, support deviation, and center of gravity approach increase synchronously, the candidate violation behavior type is determined to be unbalanced tilting behavior. The candidate violation behavior type, personnel identification, and frame number data are correlated to generate candidate violation behavior data.

[0029] In this embodiment, obtaining the results of identifying violations by high-altitude workers specifically includes: Perform time-series continuous verification on candidate violation data to determine valid time-series candidate violation data; The process for determining time-series valid candidate violation data is as follows: Candidate violation data are aggregated according to personnel identification and candidate violation type to obtain the same candidate violation type corresponding to the same worker; the aggregated candidate violation data are arranged in chronological order according to frame number data, and the number of consecutive frames of the same candidate violation type in continuous work site image frames is counted; the number of consecutive frames is compared with a preset consecutive frame threshold, and when the number of consecutive frames reaches the preset consecutive frame threshold, the corresponding candidate violation data is determined as time-series valid candidate violation data. Based on the time-series valid candidate violation data, the frame number data of the same personnel identification and the same candidate violation type in the continuous operation site image frames are read, and the attitude stability is verified according to the maintenance state of the candidate violation type and the continuous state of the frame number data in the continuous operation site image frames to generate stable candidate violation data. The specific process for generating stable candidate violation data is as follows: Based on personnel identification and candidate violation type, the same candidate violation type corresponding to the same worker is aggregated, and the aggregated time-series valid candidate violation data is arranged chronologically according to frame number data; it is determined whether the frame number data in adjacent work site image frames is continuous, and whether the candidate violation type in adjacent work site image frames is consistent; when the same candidate violation type corresponding to the same personnel identification remains consistent in consecutive work site image frames, and the corresponding frame number data remains continuous, the time-series valid candidate violation data is determined as stable candidate violation data; when the candidate violation type changes in consecutive work site image frames, the corresponding time-series valid candidate violation data is removed from the stable candidate violation data. The time-series valid candidate violation data that maintains the same candidate violation type and has continuous frame number data in the stable candidate violation data are identified as valid violation data, and the violation identification results of high-altitude workers are obtained. The violation identification results of high-altitude workers include the valid violation type, personnel identification, start frame number, end frame number and duration frame number.

[0030] Example 1: In a curtain wall installation project for a commercial complex in a certain city, the construction unit needed to erect a suspended platform on the outside of the building from the 12th to the 15th floors. Workers inside the platform were responsible for positioning the curtain wall glass, cleaning sealant, and installing fasteners. In this scenario, the suspended platform was narrow, requiring workers to frequently turn around, reach out, and bend over to retrieve tools, and there was a significant risk of near-edge contact on the outside of the platform. Traditional monitoring methods relied mainly on ground safety officers patrolling and manual review of video footage in the monitoring room. However, due to the high position of the suspended platform and the rapid changes in personnel movement, safety officers found it difficult to promptly detect dangerous behaviors such as workers leaning over the railing, abnormal footsteps, or shifting their center of gravity outwards. When workers briefly leaned over to retrieve materials or crossed the railing to adjust components, ordinary video surveillance could only record the scene and could not accurately determine whether the action constituted a violation.

[0031] High-definition and depth cameras are deployed outside the high-altitude work area to continuously capture images of the suspended platform, guardrails, edge areas, and personnel movement areas, and to collect work space calibration data. The system first preprocesses the image data to generate a sequence of high-altitude work image frames. Then, based on the work space calibration data, it reconstructs the scene baseline for the suspended platform, guardrail boundaries, edge areas, and personnel movement areas, generating high-altitude work scene baseline data. This high-altitude work scene baseline data serves as the reference basis for subsequent judgments regarding whether personnel posture has crossed boundaries, approached edges, or deviated from the safe activity area.

[0032] In field applications, the system performs human contour segmentation and personnel region anchoring on the image frame sequence of high-altitude operations to obtain a personnel posture observation sequence. For the same worker, the system continuously associates their personnel region anchoring box according to the frame number data to avoid identity confusion when multiple people are working simultaneously. Subsequently, the system performs human posture capture on the personnel posture observation sequence to obtain a sequence of worker posture key points, and then refines the posture trajectory of the worker posture key point sequence to form personnel posture trajectory data. The personnel posture trajectory data includes torso orientation trajectory, limb extension trajectory, foot support trajectory, and human center of gravity trajectory, which can reflect the movement changes of the worker in continuous on-site image frames.

[0033] In further identification, the system performs attitude-boundary coupling projection on the personnel posture trajectory data and the high-altitude operation scenario reference data to obtain attitude-boundary coupling data. Through this processing, the system can determine whether the worker's torso orientation continuously converges towards the edge reference area, whether the limb extension trajectory crosses the guardrail reference line, whether the foot support trajectory deviates from the personnel's movable reference area, and whether the body's center of gravity trajectory is close to the edge reference area. Subsequently, the attitude-boundary coupling data is input into the improved CleverCatch model, which uses attitude-boundary coupling shaping, edge posture risk focusing, violation posture link tracking, and candidate violation behavior extraction processing to generate candidate violation behavior data. For actions such as brief turns, normal reaching for installation, and bending over to pick up tools, the system filters them through temporal continuity verification and attitude stability verification; for actions such as continuously leaning out of the guardrail, foot support deviating from the safe area, and the center of gravity continuously approaching the edge area, the system identifies them as valid violation behavior data and obtains the high-altitude worker violation behavior identification results.

[0034] In the experiment, ten consecutive working days of suspended platform operation videos were selected as test data, with a total video duration of 48 hours, involving 18 high-altitude workers. Real violation samples were labeled through manual review, and the method of this invention was compared with ordinary manual video inspection methods and ordinary target detection and recognition methods. The manual video inspection method involves safety officers reviewing the videos and recording violations; the ordinary target detection and recognition method only identifies personnel based on bounding boxes and simple boundary crossing judgments; the method of this invention uses attitude capture, attitude trajectory refinement, attitude-boundary coupled projection, and an improved CleverCatch model for recognition. Test results show that the method of this invention can more accurately identify continuous near-edge dangerous actions and can reduce false alarms caused by normal reaching, turning, and short-term occlusion.

[0035] Table 1 Comparison of the effectiveness of identifying violations by high-altitude workers

[0036] As shown in Table 1, while manual video surveillance can identify violations based on on-site experience, it suffers from numerous missed detections due to factors such as safety officer attention, screen obstruction, and multiple personnel working simultaneously. The average recognition delay reaches 18.6 seconds, making it insufficient for real-time monitoring of high-altitude operations. While conventional target detection and recognition methods can quickly detect personnel approaching dangerous areas, they rely primarily on personnel detection frames and cannot accurately analyze torso orientation, limb extension, foot support, and center of gravity trajectories. Therefore, they are prone to false alarms in normal scenarios such as reaching for installations or bending to retrieve tools.

[0037] The method of this invention achieved 58 correct identifications, with an accuracy rate of 93.55%, reducing the number of missed detections to 4, the number of false alarms to 5, and the average identification delay to 2.4 seconds. These results demonstrate that this invention provides spatial reference through high-altitude work scenario baseline data, reflects changes in worker movements through personnel posture trajectory data, and analyzes the relationship between worker posture and the adjacent area, guardrail baseline, and worker's movable area through posture boundary coupling data and an improved CleverCatch model. This enables a more stable distinction between normal work actions and illegal or dangerous actions. Compared to existing methods, this invention improves the accuracy, timeliness, and stability of identifying violations in high-altitude work, reduces omissions during manual inspections, and provides more reliable technical support for on-site safety supervision.

[0038] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying violations by high-altitude workers based on attitude capture, characterized in that, Includes the following steps: Collect on-site image data and work space calibration data of the high-altitude work area, and preprocess the on-site image data to generate a high-altitude work image frame sequence; Based on the workspace calibration data, the high-altitude work area is reconstructed to generate high-altitude work scene benchmark data. Based on the high-altitude operation image frame sequence, human body contour segmentation and personnel region anchoring are performed on the workers to generate a personnel posture observation sequence. Human posture is captured from the personnel posture observation sequence to generate a sequence of worker posture key points; The posture trajectory of the operator's posture key point sequence is refined to generate the operator's posture trajectory data; The attitude-boundary coupled projection is performed between the personnel attitude trajectory data and the high-altitude operation scenario baseline data to obtain attitude-boundary coupled data. The attitude boundary coupling data is input into the improved CleverCatch model to generate candidate violation data. The candidate violation data are subjected to time-series continuity verification and attitude stability verification to generate valid violation data and obtain the violation identification results of high-altitude workers.

2. The method for identifying violations by high-altitude workers based on attitude capture according to claim 1, characterized in that, Image acquisition equipment is deployed in the high-altitude work area to continuously capture images of the work platform, guardrails, edge areas, personnel activity areas, and high-altitude workers, thereby obtaining image data of the work site and acquiring work space calibration data corresponding to the high-altitude work area. The preprocessing includes frame sequence extraction, invalid frame removal, image denoising, distortion correction, illumination equalization, and timestamp synchronization processing.

3. The method for identifying violations by high-altitude workers based on attitude capture according to claim 1, characterized in that, The generation of the high-altitude operation scenario reference data specifically includes: The coordinate system of the workspace calibration data is processed to generate spatial boundary calibration data. Image coordinate transformation is performed on the spatial boundary calibration data to generate image boundary calibration data corresponding to the high-altitude operation image frame sequence; The image boundary calibration data is parsed for boundary type and reconstructed for boundary morphology to generate a set of scene reference units; Based on the positional relationships between the scene reference units in the scene reference unit set, a scene reference structure is constructed; By associating the scene baseline structure with the image boundary calibration data, high-altitude operation scene baseline data is generated.

4. The method for identifying violations by high-altitude workers based on attitude capture according to claim 1, characterized in that, The generation of the personnel posture observation sequence specifically includes: Configure frame numbers for each work site image frame in the high-altitude operation image frame sequence to generate frame number data; Based on frame number data, personnel foreground is extracted from each work site image frame in the high-altitude operation image frame sequence to generate personnel pixel regions. Edge convergence and hole filling processes are applied to the pixel regions of people to generate human body contour segmentation data. The contour boundaries of the human body contour segmentation data are extracted to generate personnel region anchor boxes, and the personnel region anchor boxes are associated with frame number data to generate worker region anchor data. The anchor boxes of personnel regions in adjacent work site image frames are continuously matched, personnel identity identifiers are assigned to the continuously matched personnel region anchor boxes, and the personnel region anchor data are time-series collected to generate a personnel posture observation sequence.

5. The method for identifying violations by high-altitude workers based on attitude capture according to claim 1, characterized in that, The generation of the operator's posture key point sequence specifically includes: Based on the personnel posture observation sequence, personnel posture image regions are extracted to generate a personnel posture image region sequence; The scale normalization process is performed on the personnel pose image region sequence, and the correspondence between the region coordinates before and after the scale normalization process is recorded to generate pose capture input data and region coordinate reconstruction data. The posture capture input data is input into the human posture capture network for thermal response extraction, generating joint thermal response data. Peak positions are extracted and confidence levels are filtered for joint thermal response data. The normalized coordinates of each human joint in the posture capture input data are determined. Combined with the regional coordinate restoration data, the normalized coordinates of the joints are restored to the image coordinate system of the corresponding work site image frame to generate single-frame posture joint data. The posture key point data of the same worker in continuous work site image frames are arranged in time sequence to generate the worker posture key point sequence.

6. The method for identifying violations by high-altitude workers based on attitude capture according to claim 1, characterized in that, The generation of the personnel posture trajectory data specifically includes: Arrange the single-frame attitude joint data in the operator's attitude joint sequence in a temporal order to generate joint timing data; The image coordinates of the same human joint in the time series data of joint points are concatenated across frames in continuous work site image frames to generate joint point motion trajectory data. Based on the motion trajectory data of the joints, the changes in torso orientation, limb extension, foot support, and center of gravity of the same worker in continuous work site image frames are extracted to generate posture state feature data. The posture trajectory is refined from the posture state feature data to generate personnel posture trajectory data.

7. The method for identifying violations by high-altitude workers based on attitude capture according to claim 1, characterized in that, The acquisition of the attitude boundary coupling data specifically includes: The posture trajectory data of the same worker in the same work site image frame is matched with the reference data of the high-altitude work scene to generate posture scene matching data. Generate posture projection element data based on posture scene matching data; Generate boundary projection reference data based on posture scene matching data; The attitude projection element data is projected onto the image coordinate system corresponding to the boundary projection reference data to generate attitude boundary position data and form attitude boundary coupling data.

8. The method for identifying violations by high-altitude workers based on attitude capture according to claim 1, characterized in that, The generation of the candidate violation data specifically includes: The improved CleverCatch model incorporates attitude boundary coupling data as input. This improved model includes an attitude boundary coupling shaping module, a near-edge attitude risk focusing module, a violation attitude link tracking module, and a candidate violation behavior extraction module. The improvements are as follows: the traditional CleverCatch model captures and classifies changes in the target state in the input data, focusing on identifying changes in target behavior in continuous data. The improved CleverCatch model uses attitude boundary coupling data as input, performing coupling shaping on the positional relationships between the torso orientation trajectory, limb extension trajectory, foot support trajectory, human center of gravity trajectory, and the scene's baseline structure. It also introduces an attitude risk vortex traction mechanism, generating near-edge attitude risk features based on the direction, magnitude, and duration of changes in orientation offset, extension overshoot, support deviation, and center of gravity approach. Furthermore, it performs cross-frame risk vortex residual tracking to generate violation attitude link data, forming candidate violation behavior data. In the posture boundary coupling shaping module, the positional relationship between the torso orientation trajectory, limb extension trajectory, foot support trajectory, human body center trajectory and scene reference structure in the posture boundary coupling data is coupled and shaped to generate posture boundary coupling features. In the edge posture risk focusing module, a posture risk eddy traction mechanism is introduced. Based on the posture boundary coupling characteristics, posture risk eddy response data is constructed. Based on the posture risk eddy response data, risk focusing processing is performed on posture changes that continuously converge toward the edge reference area, limb changes that continuously cross the guardrail reference line, foot support changes that continuously deviate from the personnel's movable reference area, and changes in the human body's center of gravity that continuously approach the edge reference area, thereby generating edge posture risk characteristics. In the violation posture link tracking module, cross-frame risk eddy residual tracking processing is performed on the edge posture risk characteristics. According to the frame number data, the continuous edge posture risk characteristics corresponding to the same personnel identification are read to generate violation posture link nodes. Based on the start frame, continuation frame and fallback frame of the violation posture link nodes in the continuous work site image frames, the edge posture risk change process of the same worker is linked together to generate violation posture link data. In the candidate violation extraction module, candidate violation types are extracted based on the violation posture link data, and candidate violation types, personnel identification, and frame number data are associated to generate candidate violation data.

9. The method for identifying violations by high-altitude workers based on attitude capture according to claim 1, characterized in that, The specific methods for obtaining the identification results of violations by high-altitude workers include: Perform time-series continuous verification on candidate violation data to determine valid time-series candidate violation data; Based on the time-series valid candidate violation data, the frame number data of the same personnel identification and the same candidate violation type in the continuous operation site image frames are read, and the attitude stability is verified according to the maintenance state of the candidate violation type and the continuous state of the frame number data in the continuous operation site image frames to generate stable candidate violation data. The time-series valid candidate violation data that maintains the same candidate violation type and has continuous frame number data in the stable candidate violation data are identified as valid violation data, thus obtaining the violation identification results of high-altitude workers.

Images