Power transmission project construction site safety violation intelligent identification and closed-loop disposal system

Abstract

This invention discloses an intelligent identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites, belonging to the field of computer vision and intelligent video image analysis. It includes acquiring PTZ camera video streams, engineering coordinates, and work plans to construct a dataset; performing multi-target tracking on video frames, calculating the cross-frame dynamic distance between personnel displacement and boundaries, and constructing a temporal behavior directed graph; combining safety regulation topology and time windows to extract sub-graph paths deviating from safety regulations and outputting initial violation judgments; parsing spatial locations and linking adjacent PTZ cameras to acquire supplementary videos, recalculating probability distributions to generate supplementary evidence collection results; extracting violation video frames and encapsulating them with time limits to generate alarm dispatch orders; receiving rectification feedback to drive PTZ cameras to acquire review videos, calculating the defect elimination fitting degree, and outputting archived logs. This invention employs multi-source visual spatiotemporal graph comparison and multi-PTZ camera collaborative dynamic evidence collection, effectively overcoming occlusion interference and improving the accuracy of violation identification and the automation level of closed-loop handling.

Description

Technical Field

[0001] This invention relates to the field of computer vision and intelligent video image analysis, and in particular to an intelligent identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites. Background Technology

[0002] The construction sites for power transmission and transformation projects are complex environments with numerous high-risk operations. To ensure inherent safety during construction, it is typically necessary to utilize computer vision and video image understanding technologies to intelligently analyze the monitoring video streams at the construction site, enabling the monitoring of unsafe behaviors and potential safety hazards related to violations by on-site personnel.

[0003] In related technologies, Chinese invention patent application CN115205674A discloses an intelligent identification system for quality defects in power transmission and transformation projects, including: a sample data acquisition module; an image conversion and filtering intelligent judgment algorithm construction module; a background extraction model and intelligent analysis algorithm construction module; a model construction module; a model training module; and a power transmission and transformation project defect intelligent identification and diagnosis model construction module. This system constructs a deep learning image recognition and diagnosis model to mine and analyze on-site video and image data of power transmission and transformation projects collected by drones and surveillance cameras, thereby automatically identifying on-site quality defects.

[0004] The aforementioned technologies primarily rely on image appearance features for static defect classification and detection, making it difficult to accurately capture the continuous dynamic behavioral changes of construction workers in complex work scenarios. This results in low accuracy in identifying dynamic violations, leading to frequent false alarms and missed detections. Furthermore, the system only provides one-way identification of abnormal states and cannot automate visual verification of the rectification and elimination status of potential violations through video images. Subsequent on-site confirmation of compliance still heavily relies on manual intervention, resulting in low efficiency in the closed-loop handling of safety violations. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an intelligent identification and closed-loop handling system for safety violations at power transmission and transformation construction sites. It employs multi-source visual spatiotemporal image comparison and multi-PTZ camera collaborative dynamic evidence collection, which can effectively overcome occlusion interference and improve the accuracy of violation identification and the level of automation in closed-loop handling.

[0006] The above objectives can be achieved through the following approach: A smart identification and closed-loop handling system for safety violations at power transmission and transformation construction sites includes a construction scene dataset construction module, used to acquire PTZ camera video streams, engineering tower coordinate data, work plan data, and violation knowledge base data. This module performs frame extraction on the PTZ camera video streams to extract video frame sequences, and analyzes the engineering tower coordinate data and work plan data to extract work area boundary coordinates and work time windows to construct the construction scene dataset. A temporal behavior directed graph construction module is used to extract video frame sequences and work area boundary coordinates from the construction scene dataset, perform multi-target tracking on the video frame sequences, calculate the cross-frame dynamic distance between the personnel node displacement vector and the work area boundary coordinates, and construct a temporal behavior directed graph using the cross-frame dynamic distance as an edge attribute. An initial violation judgment module is used to extract the safety regulation condition topology from the violation knowledge base data, combine it with the work time window, perform distance calculation on the temporal behavior directed graph, and extract subgraph paths that deviate from the safety regulation condition topology. The system outputs an initial violation judgment result; a PTZ camera supplementary evidence collection module, used to parse the three-dimensional spatial position based on the initial violation judgment result, generate a field-of-view adjustment command and send it to the adjacent collaborative PTZ cameras at the three-dimensional spatial position, fuse the return supplementary video frames from the adjacent collaborative PTZ cameras to calculate the joint probability distribution, and generate a supplementary evidence collection judgment result; a violation alarm dispatch module, used to map the violation type tag of the supplementary evidence collection judgment result to the violation knowledge base data to extract the rectification time limit parameter, extract the violation video frame segment of the supplementary evidence collection judgment result, encapsulate the violation video frame segment, the rectification time limit parameter and the work area boundary coordinates, and generate violation alarm dispatch data; a rectification review and archiving module, used to receive rectification feedback instructions for the violation alarm dispatch data, drive the associated PTZ cameras at the work area boundary coordinates to obtain the review video stream, calculate the defect fitting degree between the visual features of the review video stream and the supplementary evidence collection judgment result, and output a closed-loop archive log based on the defect fitting degree.

[0007] Optionally, the construction scene dataset construction module includes: a video feature calculation unit, used to parse the temporal coding structure of the PTZ camera video stream, perform dynamic keyframe extraction, strip static background images, and generate a video frame sequence; and a multi-dimensional spatiotemporal fusion unit, used to extract the three-dimensional geodetic coordinate values ​​of the engineering tower coordinate data, combine the scheduling field in the work plan data to perform polygon bounding box spatial mapping, generate the boundary coordinates of the work area and the work time window, and stitch the video frame sequence to construct the construction scene dataset.

[0008] Optionally, the temporal behavior directed graph construction module includes: an entity trajectory calculation unit, used to extract video frame sequences from the construction scene dataset, perform pixel-level instance segmentation and cross-frame feature point matching on the video frame sequences, extract the centroid coordinates of the target bounding boxes in consecutive frames, and calculate the temporal difference of the centroid coordinates to generate personnel node displacement vectors; and a topology graph construction unit, used to extract the boundary coordinates of the work area in the construction scene dataset, calculate the shortest spatial distance between the personnel node displacement vector and the boundary coordinates of the work area, generate cross-frame dynamic distance, extract the personnel node displacement vector as a graph node attribute, map the cross-frame dynamic distance as edge connection weights, and construct a temporal behavior directed graph.

[0009] Optionally, the initial violation determination module includes: a spatiotemporal graph mapping unit, used to extract the safety regulation condition topology of the violation knowledge base data and the operation time window of the construction scenario dataset, perform node timestamp embedding mapping from the operation time window to the safety regulation condition topology, and construct a benchmark constraint graph; and a graph edit distance calculation unit, used to perform graph edit distance calculation on the temporal behavior directed graph and the benchmark constraint graph, filter abnormal node sequences, extract the connected topology structure of the abnormal node sequences as subgraph paths deviating from the safety regulation condition topology, and output the initial violation determination result.

[0010] Optionally, the PTZ camera supplementary evidence collection module includes: a spatial geometric perspective positioning unit, used to extract the two-dimensional pixel coordinate matrix of the initial violation judgment result, perform perspective projection transformation of the two-dimensional pixel coordinate matrix to the world coordinate system, resolve the three-dimensional spatial position, calculate the PTZ rotation vector and zoom factor of the three-dimensional spatial position and the adjacent collaborative PTZ camera, generate a field of view adjustment command and issue it; and a cross-viewpoint tensor fusion recalculation unit, used to receive the supplementary video frames returned by the adjacent collaborative PTZ camera based on the field of view adjustment command, extract the supplementary visual feature tensor of the returned supplementary video frames, align and stitch the supplementary visual feature tensor with the initial visual feature execution channel of the sub-graph path, calculate the joint probability distribution, and generate a supplementary evidence collection judgment result.

[0011] Optionally, the system further includes: expanding the spatial range of the boundary coordinates of the work area based on the frequency of violations in the initial violation determination result, and generating a dynamic warning and defense line level for the corresponding construction area.

[0012] Optionally, the violation alarm dispatch module includes: a time-series interval interception unit, used to parse the violation type tag of the supplementary evidence collection judgment result, perform entity index matching between the violation type tag and the violation knowledge base data, extract rectification time limit parameters, and parse the start timestamp and end timestamp of the violation type tag, intercept continuous video images from the start timestamp to the end timestamp in the supplementary evidence collection judgment result, and generate violation video frame segments; and a dynamic routing encapsulation unit, used to extract the boundary coordinates of the work area and the dynamic warning defense level, convert the dynamic warning defense level into dispatch routing weights, and perform markup language serialization encoding on the violation video frame segments, the rectification time limit parameters and the boundary coordinates of the work area, fuse the dispatch routing weights, and generate violation alarm dispatch data.

[0013] Optionally, the rectification review and archiving module includes: a review drive and field of view calibration unit, used to parse the trigger timestamp of the rectification feedback instruction, map the boundary coordinates of the work area to a PTZ directional control matrix, and drive the associated PTZ camera to obtain the review video stream based on the PTZ directional control matrix at the trigger timestamp; and a cross-period feature comparison and calculation unit, used to extract the current visual feature vector of the review video stream, extract the benchmark violation feature vector of the supplementary evidence judgment result, calculate the spatial difference scalar between the current visual feature vector and the benchmark violation feature vector, generate a defect elimination fitting degree, encapsulate it into a structured data block, and output a closed-loop archived log.

[0014] Based on the same inventive concept, this invention also provides a method for intelligent identification and closed-loop handling of safety violations at power transmission and transformation engineering construction sites. The method includes: acquiring PTZ camera video streams, engineering tower coordinate data, work plan data, and violation knowledge base data; performing frame extraction on the PTZ camera video streams to extract video frame sequences; parsing the engineering tower coordinate data and the work plan data to extract work area boundary coordinates and work time windows to construct a construction scene dataset; using the construction scene dataset to extract video frame sequences and work area boundary coordinates; performing multi-target tracking on the video frame sequences; calculating the cross-frame dynamic distance between the personnel node displacement vector and the work area boundary coordinates; constructing a time-series behavior directed graph using the cross-frame dynamic distance as an edge attribute; extracting the safety regulation condition topology from the violation knowledge base data; combining it with the work time window; performing distance calculation on the time-series behavior directed graph; and extracting deviations from the safety regulation conditions. The subgraph path of the topology is used to output the initial violation judgment result; based on the initial violation judgment result, the three-dimensional spatial position is parsed, and a field of view adjustment command is generated and sent to the adjacent cooperative PTZ camera at the three-dimensional spatial position. The joint probability distribution is calculated by fusing the supplementary video frames returned by the adjacent cooperative PTZ camera to generate a supplementary evidence collection judgment result; according to the violation type label of the supplementary evidence collection judgment result, the rectification time limit parameter is extracted by mapping the violation knowledge base data, the violation video frame segment of the supplementary evidence collection judgment result is extracted, and the violation video frame segment, the rectification time limit parameter and the boundary coordinate of the work area are encapsulated to generate violation alarm dispatch data; the rectification feedback command for the violation alarm dispatch data is received, and the associated PTZ camera at the boundary coordinate of the work area is driven to obtain the review video stream. The visual features of the review video stream and the defect elimination fitting degree of the supplementary evidence collection judgment result are calculated, and the closed-loop archived log is output based on the defect elimination fitting degree.

[0015] Compared with the prior art, the present invention has the following advantages: 1. Upgrade isolated target detection to scenario-based dynamic temporal analysis, which removes the confusing features between legitimate construction operations and real violations, and reduces the false alarm rate caused by complex backgrounds or similar actions; 2. By jointly calculating the probability distribution across different perspectives, active cross-validation of violations in occluded environments or low-confidence scenarios is achieved, improving recognition accuracy and evidence completeness. 3. It eliminates the problem of inconsistent standards caused by human intervention, which not only shortens the processing cycle, but also ensures the authenticity and immutability of security data.

[0016] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a framework diagram of the intelligent identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the intelligent identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites according to an embodiment of the present invention; Figure 3 This is a data distribution diagram of the distance cost of image editing under different behavior modes in an embodiment of the present invention; Figure 4 This is a diagram showing the evolution of the violation confidence probability before and after cross-view tensor fusion in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Reference Figure 1 One embodiment of the present invention proposes an intelligent identification and closed-loop handling system for safety violations at power transmission and transformation construction sites. It adopts multi-source visual spatiotemporal image comparison and multi-PTZ camera collaborative dynamic evidence collection, which can effectively overcome occlusion interference and improve the accuracy of violation identification and the level of automation of closed-loop handling.

[0021] like Figure 2 As shown, the system in this embodiment specifically includes: The construction scenario dataset construction module is used to acquire PTZ camera video stream, engineering tower coordinate data, work plan data and violation knowledge base data, perform frame extraction operation on the PTZ camera video stream to extract video frame sequences, and parse the engineering tower coordinate data and the work plan data to extract the work area boundary coordinates and work time window to construct the construction scenario dataset. Optionally, the construction scenario dataset construction module includes: The video feature processing unit is used to parse the temporal coding structure of the PTZ camera video stream, perform dynamic keyframe extraction, strip static background images, and generate a video frame sequence. The video feature calculation unit acquires the PTZ camera video stream transmitted by the front-end acquisition device and parses the underlying communication encapsulation protocol of the PTZ camera video stream to read the timing coding structure. The timing coding structure contains continuous video frame data arranged in absolute time. In order to remove static background images and accurately retain the monitoring screen containing the movement of people or equipment, the video feature calculation unit performs dynamic keyframe extraction based on adaptive background updates on the continuous video frames. The video feature calculation unit extracts the pixel feature matrix of the current frame and the background pixel matrix of the previous moment, and calculates the sum of the absolute differences between the pixel feature matrix of the current frame and the background pixel matrix of the previous moment. In order to filter out visual noise interference caused by gradual changes in natural lighting and slight mechanical shaking of the camera, an environmental noise benchmark threshold is introduced. The value of this environmental noise benchmark threshold is not determined by subjective experience, but is obtained by collecting 500 hours of static video recording data of the corresponding model of PTZ camera in an idle state, extracting the maximum pixel fluctuation variance and calculating it through scalar amplification, so as to ensure the objectivity of the threshold and the physical support basis. When the sum of the calculated absolute differences is greater than the environmental noise baseline threshold, the current frame is determined to have significant dynamic behavior features, and this current frame is extracted as a dynamic keyframe and stored in the video frame sequence. When the sum of the absolute differences is less than or equal to the environmental noise baseline threshold, the current frame is determined to be a static background image and is discarded. Simultaneously, to adapt to gradual changes in outdoor weather, the video feature calculation unit uses a dynamic learning rate parameter to perform smooth update calculations on the background model. The update calculation relationship satisfies the following formula: , Among them, letters Represents the current background pixel matrix generated after the update calculation; letters Represents the pixel feature matrix of the current image; letters Represents the background pixel matrix from the previous time step, read from the cache; letters This represents the dynamic learning rate parameter. The physical meaning of this dynamic learning rate parameter is to control the speed at which the background image absorbs new pixel features. Its value is calculated using a sinusoidal periodic function of the ambient lighting changes and fixed as a small decimal. This ensures that the background model can slowly integrate changes in cloud shadows, preventing the misjudgment and integration of pixels representing rapidly moving construction workers into the background model.

[0022] For example, the video feature calculation unit receives a real-time PTZ camera video stream with a resolution of 1920 x 1080. Through analysis of the device initialization configuration, the environmental noise baseline threshold for the construction site is determined to be 15000, and the dynamic learning rate parameter is set to 0.05. At the 10th second of the time sequence, the video feature calculation unit extracts the pixel feature matrix of the current frame and calculates the sum of absolute differences pixel by pixel with the background pixel matrix from the previous moment. The calculation shows that due to the presence of a large insulated bucket truck entering the construction area within the monitoring field of view, the pixel array changes drastically, resulting in an output absolute difference sum of 450000. This output value is much greater than the environmental noise baseline threshold of 15000. Therefore, the video feature calculation unit explicitly marks this monitoring frame as a dynamic keyframe and loads it into the video frame sequence buffer. During the subsequent period from the 11th to the 15th second, since the insulated bucket truck stopped moving and no personnel were on site, the sum of the absolute differences calculated over these five consecutive seconds fluctuated between 8000 and 12000, all strictly below the environmental noise baseline threshold. The video feature calculation unit determined this continuous frame sequence as a static background image without effective activity and directly removed and cleaned it. Simultaneously, substituting the background update formula, and using a dynamic learning rate parameter of 0.05, the subtle changes in light in the current image were slowly integrated into the historical background with a weight of 5%, outputting an updated current background pixel matrix. This effectively removed the spatiotemporal redundancy of the video stream, generating a high-information-density video frame sequence.

[0023] The multi-dimensional spatiotemporal fusion unit is used to extract the three-dimensional geodetic coordinate values ​​of the engineering tower coordinate data, perform polygon bounding box spatial mapping in combination with the scheduling field in the work plan data, generate the boundary coordinates of the work area and the work time window, and stitch the video frame sequence to construct the construction scene dataset.

[0024] The multi-dimensional spatiotemporal fusion unit receives location data packets sent by the surveying terminal and extracts the three-dimensional geodetic coordinates from the engineering tower coordinate data. These three-dimensional geodetic coordinates include the absolute longitude and absolute latitude values ​​of each infrastructure tower node. To transform the discrete tower geographic coordinates into a geofence covering the actual physical work area, the multi-dimensional spatiotemporal fusion unit performs polygon bounding box spatial mapping. This spatial mapping mechanism does not generate simple point connections but fully considers the turning radius of large construction machinery and the impact range of falls from heights, performing spatial buffer expansion calculations based on the three-dimensional geodetic coordinates. A safety buffer distance parameter is introduced, the value of which is strictly derived from the maximum voltage level safe discharge distance explicitly specified in the power industry's power transmission and transformation engineering safety regulations, and is the sum of the values ​​of the maximum working arm length of common tower crane equipment. The multi-dimensional spatiotemporal fusion unit traverses and compares the three-dimensional geodetic coordinates of all tower nodes, extracts the maximum longitude, minimum longitude, maximum latitude, and minimum latitude values, and performs topology expansion using a safety buffer distance parameter to calculate the longitude relationship of the boundary coordinates of the work area, satisfying the following formula: , , Among them, letters Represents the minimum longitude boundary value of the work area boundary coordinates after the expansion calculation; letter Represents the maximum longitude boundary value of the work area boundary coordinates after the expansion calculation; letter Represents the minimum longitude value among all extracted 3D geodetic coordinates; the letter... Represents the maximum longitude value among all extracted 3D geodetic coordinates; the letter... This represents the safety buffer distance parameter after converting the physical distance measured in meters into angular offsets based on the Earth's curvature. The latitudinal boundary values ​​of the work area boundary coordinates are calculated using the same addition and subtraction offset algorithm logic. Through the above rigorous geometric topological mapping, the work area boundary coordinates, including the coordinate parameters of the four extreme boundaries, are generated. Simultaneously, the multidimensional spatiotemporal fusion unit parses the scheduling field format in the work plan data, extracts the physical timestamp of obtaining the construction permit on the current day and the timestamp of clearing and leaving the site, and combines these two endpoints to generate a one-dimensional work time window. Finally, the multidimensional spatiotemporal fusion unit uses the video frame sequence generated in the same physical space as the time series feature data, the work area boundary coordinates as the spatial geometric constraint, and the work time window as the temporal life cycle constraint. This three-dimensional data volume is spliced ​​and stored in the form of key-value pairs to construct a construction scene dataset with strict physical spatiotemporal constraint mapping relationships.

[0025] For example, the multi-dimensional spatiotemporal fusion unit extracted the three-dimensional geodetic coordinates of the main towers of a substation expansion project from the surveying documents. The resolved latitude and longitude coordinate sets are as follows: first node: longitude 118.3200 degrees, latitude 31.4500 degrees; second node: longitude 118.3200 degrees, latitude 31.4550 degrees; third node: longitude 118.3250 degrees, latitude 31.4500 degrees. According to relevant industry safety operation specifications, the required safe operating radius for this high-voltage work area needs to be extended outward by 50 meters. The multi-dimensional spatiotemporal fusion unit used the Earth radius coefficient to calculate and convert the 50-meter physical safety buffer distance parameter into an equivalent latitude and longitude offset, with a calculated output of approximately 0.0005 degrees. The multi-dimensional spatiotemporal fusion unit substitutes the numerical values ​​into the polygon bounding box spatial mapping formula, extracts the minimum longitude value of 118.3200 degrees from the three nodes, subtracts the offset of 0.0005 degrees, and obtains the minimum longitude boundary value of 118.3195 degrees; it also extracts the maximum longitude value of 118.3250 degrees, adds the offset of 0.0005 degrees, and obtains the maximum longitude boundary value of 118.3255 degrees. Using the same mathematical logic, the minimum latitude boundary value of 31.4495 degrees and the maximum latitude boundary value of 31.4555 degrees are calculated respectively. These four expanded limit values ​​together form a rectangular vector geofence, thus establishing the boundary coordinates of the work area. At the same time, the multi-dimensional spatiotemporal fusion unit reads the work plan data of the work order issued on the same day, parses the scheduling field content, extracts the physical timestamp of entry as 08:00:00 on the same day, and the timestamp of clearing and leaving the site as 18:00:00 on the same day, and generates the work time window within this interval. Subsequently, the multi-dimensional spatiotemporal fusion unit will be timestamped with all video frame sequences collected between 08:00:00 and 18:00:00, and then attached and spliced ​​to the nodes of the rectangular vector geofence and the operation time window mentioned above, ultimately generating a complete construction scene dataset bottom-level record containing multimodal spatiotemporal constraint information.

[0026] The temporal behavior directed graph construction module is used to extract video frame sequences and work area boundary coordinates using the construction scene dataset, perform multi-target tracking on the video frame sequences, calculate the cross-frame dynamic distance between the personnel node displacement vector and the work area boundary coordinates, and construct a temporal behavior directed graph with the cross-frame dynamic distance as the edge attribute. Optionally, the directed graph construction module for temporal behavior includes: The entity trajectory calculation unit is used to extract the video frame sequence of the construction scene dataset, perform pixel-level instance segmentation and cross-frame feature point matching on the video frame sequence, extract the centroid coordinates of the target bounding boxes of consecutive frames, and calculate the temporal difference of the centroid coordinates to generate personnel node displacement vectors. The entity trajectory solving unit extracts a sequence of video frames tagged with timestamps from the construction scene dataset. To accurately track each worker in complex power transmission and transformation construction scenarios, the entity trajectory solving unit performs pixel-level instance segmentation on each frame in the video frame sequence. Pixel-level instance segmentation generates an independent set of pixel masks and corresponding target bounding boxes for each identified individual in the image space by calling the underlying convolutional neural network feature extraction layer, thereby solving the bounding box confusion problem caused by multiple overlapping or equipment occlusion. After segmentation, the entity trajectory solving unit performs cross-frame feature point matching between adjacent consecutive frames. This process extracts local texture feature vectors within each pixel mask set and calculates the cosine similarity between the feature vectors of consecutive frames. To avoid introducing subjective parameters without physical basis, the classification cutoff value for this matching judgment is derived from the data distribution backtesting of construction videos under the same lighting conditions over the past 500 hours, and the value corresponding to the extreme point of the precision curve is selected as the objective judgment standard. When the cosine similarity is greater than this judgment standard, the target bounding boxes in different frames are associated with the temporal trajectory of the same worker. For the same worker who has been successfully associated, the entity trajectory calculation unit extracts their target bounding box in the current frame, and parses the pixel row and column indices of the top-left and bottom-right extreme points of the target bounding box. The centroid coordinates of the target bounding boxes in consecutive frames are then extracted using an arithmetic mean calculation. The calculation relationship of the centroid coordinates satisfies the following formula: , , Among them, letters The horizontal pixel column index representing the centroid coordinates of the target bounding box at the current frame time; letter The vertical pixel row index representing the centroid coordinates of the target bounding box at the current frame time; letters With letters These represent the horizontal pixel column index and vertical pixel row index of the top-left corner extreme point of the target bounding box at the current frame time; the letters... With letters These represent the horizontal pixel column index and vertical pixel row index of the extreme point at the lower right corner of the target bounding box at the current frame time, respectively. Subsequently, the entity trajectory calculation unit extracts the homography perspective transformation matrix generated after the PTZ camera equipment undergoes on-site static calibration, mapping the centroid coordinates of the pixel dimension to the horizon coordinate system to generate two-dimensional physical coordinate points. The entity trajectory calculation unit extracts the two-dimensional physical coordinate points of the current frame and subtracts the two-dimensional physical coordinate points of the corresponding personnel from the previous frame. By subtracting the coordinate point values, the temporal difference of the centroid coordinates is calculated, ultimately generating a personnel node displacement vector containing the motion vector direction, motion physical amplitude, and the current terminal physical coordinate point.

[0027] For example, the entity trajectory calculation unit receives a video frame sequence extracted from a construction scene dataset. In the tenth frame, the entity trajectory calculation unit identifies an insulated worker wearing a safety helmet through pixel-level instance segmentation, and extracts the pixel coordinates of the top-left extreme point of the target bounding box as horizontal 100 and vertical 200, and the pixel coordinates of the bottom-right extreme point as horizontal 120 and vertical 240. Substituting these values ​​into the centroid coordinate calculation formula, adding 100 and 120 and dividing by 2 yields the horizontal pixel column index of the target bounding box centroid coordinates at the current frame as 110, and adding 200 and 240 and dividing by 2 yields the vertical pixel row index as 220. In the subsequent eleventh frame, the worker is successfully associated through cross-frame feature point matching, and the centroid coordinates of the eleventh frame are calculated using the same formula as horizontal 115 and vertical 225. The entity trajectory calculation unit calls the homography perspective transformation matrix to perform spatial projection, mapping the centroid coordinates of the tenth frame to two-dimensional physical coordinates in the horizontal coordinate system, namely 10.0 meters east and 15.0 meters north. The centroid coordinates of the eleventh frame are mapped to 10.5 meters east and 15.6 meters north in the horizontal coordinate system. The entity trajectory calculation unit calculates the temporal difference by subtracting numerical values: 10.5 meters minus 10.0 meters yields an eastward displacement difference of 0.5 meters, and 15.6 meters minus 15.0 meters yields a northward displacement difference of 0.6 meters. The entity trajectory calculation unit combines these values ​​to generate the personnel node displacement vector for the worker at the eleventh frame. This personnel node displacement vector contains two-dimensional vector components of 0.5 and 0.6, and simultaneously anchors the current terminal physical coordinates at 10.5 meters east and 15.6 meters north.

[0028] The topology graph construction unit is used to extract the boundary coordinates of the work area in the construction scene dataset, calculate the shortest spatial distance between the personnel node displacement vector and the boundary coordinates of the work area, generate cross-frame dynamic distance, extract the personnel node displacement vector as graph node attribute and map the cross-frame dynamic distance as edge connection weight, and construct a temporal behavior directed graph.

[0029] The topology graph construction unit extracts the pre-encapsulated boundary coordinates of the work area from the construction scene dataset. These boundary coordinates consist of multiple sequentially connected latitude and longitude boundary segments, representing the legal geofence for safe construction operations. The topology graph construction unit extracts the current terminal physical coordinates from the personnel node displacement vector and calculates the Euclidean vertical distance from this current terminal physical coordinate point to all latitude and longitude boundary segments contained within the work area boundary coordinates. The topology graph construction unit performs a minimum value traversal filtering on all calculated Euclidean vertical distances to calculate the shortest spatial distance between the personnel node displacement vector and the work area boundary coordinates. The underlying logic of this shortest spatial distance calculation relationship satisfies the following mathematical formula: , Among them, letters The displacement vector of the representative personnel node represents the current terminal physical coordinate point to the boundary coordinates of the work area. Vertical spatial distance of latitude and longitude boundary segments; letters The representative node's displacement vector represents its eastward physical coordinates in the current frame; letters The letter represents the northward physical coordinates of the personnel node displacement vector in the current frame; ,letter With letters Together they constitute the first in the boundary coordinates of the work area. The standardized constant coefficients of the latitude and longitude boundary line segments in the two-dimensional plane straight line equations are directly parsed from the boundary topology dictionary of the construction scene dataset. The minimum constant coefficients are calculated and extracted. The topology graph construction unit generates cross-frame dynamic distances. The physical meaning of cross-frame dynamic distance lies in quantitatively characterizing the degree to which a worker approaches a dangerous boundary during movement. After acquiring the above data, the topology graph construction unit instantiates an empty temporal behavior directed graph in the graph theory data structure, establishing a graph node structure for each tracked worker according to the temporal flow order of the video frame sequence. The topology graph construction unit extracts the worker node displacement vector as a graph node attribute and assigns it to the corresponding graph node to store the worker's motion state. Subsequently, the topology graph construction unit establishes directed edges representing temporal flow relationships between graph nodes in adjacent frames and maps the cross-frame dynamic distance to edge connection weights, completing the assignment of topology edge values. Through continuous iteration of the above graph node creation and directed edge connection operations, the topology graph construction unit outputs a temporal behavior directed graph containing complete temporal causal logic and spatial approximation states.

[0030] The initial violation determination module is used to extract the safety regulation condition topology of the violation knowledge base data, combine it with the operation time window, perform distance calculation on the directed graph of the time sequence behavior, extract the subgraph path that deviates from the safety regulation condition topology, and output the initial violation determination result. Optionally, the initial violation determination module includes: The spatiotemporal graph mapping unit is used to extract the safety regulation condition topology of the violation knowledge base data and the operation time window of the construction scenario dataset, perform the node timestamp embedding mapping from the operation time window to the safety regulation condition topology, and construct the benchmark constraint graph. The spatiotemporal graph mapping unit extracts the safety regulation condition topology from the violation knowledge base data. The safety regulation condition topology is a set of standard construction operation specifications encoded in a graph structure. Its internal node attributes record the upper limit benchmark value of personnel movement speed under compliant operation conditions, and its internal directed edge weights record the shortest physical isolation value between personnel and the danger zone. These upper limit benchmark values ​​and shortest physical isolation values ​​are not subjectively set manually, but are extracted from the physical restrictions on the walking speed of ground workers and the standard clearance distance for safe operation with live conductors in the National Power Industry General Transmission and Transformation Construction Safety Regulations, respectively. The spatiotemporal graph mapping unit extracts the operation time window from the construction scenario dataset. This operation time window includes the entry physical timestamp and the clearing and departure timestamp. Based on the fixed hardware sampling frequency of the front-end PTZ camera video stream, the spatiotemporal graph mapping unit calculates the sampling time interval and divides the operation time window into a continuous discrete timestamp sequence proportionally according to this sampling time interval. Subsequently, the spatiotemporal graph mapping unit performs node timestamp embedding mapping, that is, instantiating an ideal state safety regulation condition topology node at each time node of the discrete timestamp sequence. The spatiotemporal mapping unit assigns the ideal state node an absolute safety attribute of zero displacement and a constant distance from the boundary equal to the shortest physical isolation value, and topologically connects adjacent discrete timestamp nodes through directed edges. Through the above-mentioned time-dimensional serialization expansion mapping operation, the static safety condition topology is transformed into a baseline constraint map covering the entire operation time window. In the mathematical model, this baseline constraint map represents the baseline trajectory where no worker will overstep boundaries or make abnormal displacements throughout the entire legal construction cycle.

[0031] The graph editing distance calculation unit is used to perform graph editing distance calculation on the temporal behavior directed graph and the benchmark constraint graph, filter abnormal node sequences, extract the connected topology of the abnormal node sequences as subgraph paths that deviate from the topology of the safety regulations, and output the initial violation judgment result.

[0032] The graph edit distance calculation unit receives the directed graph of temporal behavior output by the directed graph construction module, and the baseline constraint graph constructed by the spatiotemporal graph mapping unit. To quantify the deviation between the actual trajectory of personnel and the safety baseline using rigorous data dimensions, the graph edit distance calculation unit performs graph edit distance calculation on the directed graph of temporal behavior and the baseline constraint graph. In the underlying graph theory algorithm, the graph edit distance represents the minimum total cost required to fit and transform the observed directed graph of temporal behavior into the baseline constraint graph through operations such as node feature replacement, structure deletion, or attribute insertion. The graph edit distance calculation unit extracts the graph node attributes of a specific target personnel in the directed graph of temporal behavior and their mapped edge connection weights, and performs cost value calculation with the baseline constraint graph features under the corresponding physical timestamp. The cost value settlement relationship satisfies the following formula: , Among them, letters The letter represents the node replacement value calculated from the associated features at the current time point; The physical magnitude of the displacement vector of a person node, obtained by parsing the graph node attributes of a directed graph of temporal behavior, represents the actual movement speed; the letter... Represents the upper limit baseline value extracted from the baseline constraint map; letters Represents the shortest physical isolation value extracted from the baseline constraint map; letter This represents the cross-frame dynamic distance extracted from the edge connection weights of a directed graph of temporal behavior. The physical meaning and mathematical mapping principle of this formula lies in the fact that the faster the actual movement speed of the worker, or the closer the actual distance to the danger boundary, the greater the risk of accidents. As the value approaches its minimum, the division ratio on the right side of the equation is amplified drastically and non-linearly, resulting in an extremely high output cost value. After obtaining the full calculation results, the graph editing distance calculation unit accumulates the node replacement cost values ​​at each timestamp along the time axis to construct an accumulated distance matrix. Subsequently, the graph editing distance calculation unit introduces an anomaly deviation judgment baseline. This anomaly deviation judgment baseline is not based on subjective presuppositions of expert experience, but is derived by extracting closed-circuit video recordings of over a thousand verified violations from the past two years of the provincial network company, performing the same graph editing distance extrapolation on the video data, and extracting the convergence extrema of the quartile interval under the normal distribution model, thus ensuring the objectivity of the judgment data. The graph editing distance calculation unit compares the accumulated distance matrix with the anomaly deviation judgment baseline, filters out continuous graph nodes whose matrix cost value is greater than the anomaly deviation judgment baseline, and extracts and generates an anomaly node sequence. Finally, the graph editing distance calculation unit extracts the connected topology of the abnormal node sequence in the original temporal behavior directed graph, uses it as a subgraph path deviating from the safety regulation condition topology, and encapsulates and outputs an initial violation judgment result containing violation type labels and violation frequency based on this topological feature. Figure 3 As shown, through kernel density estimation and discrete distribution scatter plots, it can be seen that the cost distribution of out-of-bounds violations is significantly higher than that of compliant operations. The high-dimensional convergence boundary of this distribution provides a solid data support for setting the baseline for abnormal deviation judgment.

[0033] The PTZ camera supplementary evidence collection module is used to parse the three-dimensional spatial position based on the initial violation judgment result, generate a field of view adjustment command and send it to the adjacent cooperative PTZ cameras at the three-dimensional spatial position, fuse the return supplementary video frames of the adjacent cooperative PTZ cameras to calculate the joint probability distribution, and generate a supplementary evidence collection judgment result. Optionally, the PTZ camera supplementary evidence collection module includes: The spatial geometric perspective positioning unit is used to extract the two-dimensional pixel coordinate matrix of the initial violation judgment result, perform the perspective projection transformation of the two-dimensional pixel coordinate matrix to the world coordinate system, resolve the three-dimensional spatial position, calculate the gimbal rotation vector and zoom factor of the three-dimensional spatial position and the adjacent cooperative PTZ camera, generate the field of view adjustment command and send it out. The spatial geometric perspective positioning unit extracts the initial violation judgment result output by the initial violation judgment module. This initial violation judgment result contains a sub-graph path representing the trajectory of the violation behavior. The spatial geometric perspective positioning unit extracts the two-dimensional pixel coordinates corresponding to each abnormal node from this sub-graph path and constructs a two-dimensional pixel coordinate matrix by splicing them according to the timestamp sequence. In order to restore the two-dimensional information of the image plane to the real three-dimensional space of the physical site, the spatial geometric perspective positioning unit performs a perspective projection transformation from the two-dimensional pixel coordinate matrix to the world coordinate system. This perspective projection transformation calls the intrinsic and extrinsic parameter matrices of the main control PTZ camera that captured the initial violation image. The intrinsic parameter matrix contains the focal length value and principal point coordinates obtained through factory calibration, and the extrinsic parameter matrix contains the coordinates of the three-dimensional installation center point of the main control PTZ camera in the world coordinate system determined based on engineering surveying data. Using homogeneous coordinates and the pinhole camera imaging model, the spatial geometric perspective positioning unit converts the two-dimensional pixel coordinate matrix into a three-dimensional ray vector in the camera coordinate system. A local elevation plane parameter is introduced. This local elevation plane parameter comes from the average terrain elevation attached to the engineering tower coordinate data in the construction scene dataset and serves as a constant in the ground constraint plane equation. The spatial geometric perspective positioning unit calculates the spatial intersection of the three-dimensional ray vector and the ground constraint plane equation, solving for the three-dimensional spatial position of the suspected illegal target, including absolute longitude, absolute latitude, and local elevation. After obtaining this three-dimensional spatial position, the spatial geometric perspective positioning unit searches the network topology and extracts the three-dimensional origin coordinates of the nearest adjacent cooperative PTZ camera. By calculating the spatial geometric difference matrix between the three-dimensional spatial position and the three-dimensional origin coordinates, the spatial geometric perspective positioning unit deduces the gimbal rotation vector required to align the optical axis of the adjacent cooperative PTZ camera with the target. Simultaneously, the spatial geometric perspective positioning unit calculates the zoom ratio, the calculation relationship of which satisfies the following formula: , Among them, letters Represents the calculated dimensionless zoom ratio; letters Represents the physical straight-line distance between a three-dimensional spatial location and the coordinates of the three-dimensional origin; letters The equivalent observation height of the target is uniformly set to half the average height of an adult Chinese male, i.e., 0.85 meters, to ensure that the upper body of the person fills the core area of ​​the image; letters Represents the physical vertical dimension of the PTZ camera's image sensor, derived from the hardware parameters in the device manual; letters This represents the current device's reference physical focal length. Finally, the spatial geometric perspective positioning unit encapsulates the gimbal rotation vector and zoom magnification, generates a standard protocol field-of-view adjustment command, and sends it to the target adjacent collaborative PTZ camera to perform multi-angle reshoots.

[0034] For example, the spatial geometric perspective positioning unit extracts a set of two-dimensional pixel coordinate matrices, where the two-dimensional pixel coordinates of a key anomaly node are 1200 in the horizontal column and 800 in the vertical column. The focal length of the main control PTZ camera is 1000 pixels, and the principal point coordinates are 960 in the horizontal column and 540 in the vertical column. After perspective projection transformation calculation, a three-dimensional ray vector pointing to the ground is generated. The local elevation plane parameter is taken as 0.5 meters above the reference plane. Through the intersection calculation of the ray and the plane, the three-dimensional spatial position of the illegal target at the scene is determined to be 25.0 meters east, 40.0 meters north, and 0.5 meters above the elevation. The three-dimensional origin coordinates of the retrieved adjacent cooperative PTZ camera are 10.0 meters east, 20.0 meters north, and 10.5 meters above the elevation. Using the spatial geometric difference calculation of trigonometric functions, the physical straight-line distance between the two is obtained, i.e., the parameter. The distance is 26.9 meters, and the gimbal rotation vector is calculated to be a horizontal azimuth of 56 degrees and a vertical pitch angle of -22 degrees. Substituting these values ​​into the zoom ratio calculation formula, the physical vertical dimensions of the image sensor are known. The reference physical focal length is 0.008 meters. The equivalent observation height of the target is 0.012 meters. Take 0.85 meters. Multiply the numerator 26.9 by 0.008 to get 0.2152, and the denominator 0.85 by 0.012 to get 0.0102. Dividing the numerator by the denominator, the zoom magnification is approximately 21.1x. The spatial geometric perspective positioning unit packages the azimuth angle of 56 degrees, the elevation angle of -22 degrees, and the zoom magnification of 21.1x into a field of view adjustment command, and sends it to control the adjacent collaborative PTZ camera to instantly lock onto the rear view of the suspect violating traffic rules.

[0035] The cross-viewpoint tensor fusion recalculation unit is used to receive the supplementary video frames returned by the adjacent cooperative PTZ camera based on the field of view adjustment command, extract the supplementary visual feature tensors of the returned supplementary video frames, align and splice the supplementary visual feature tensors with the initial visual feature execution channel of the subgraph path, calculate the joint probability distribution, and generate supplementary evidence collection judgment results.

[0036] The cross-view tensor fusion recalculation unit establishes a low-level streaming media monitoring service, receiving supplementary video frames passively transmitted from adjacent cooperative PTZ cameras based on field-of-view adjustment commands. To fully utilize visual evidence from different viewpoints and eliminate potential object occlusion interference in the original viewpoint, the cross-view tensor fusion recalculation unit calls a deep residual network model to perform convolutional feature extraction on the transmitted supplementary video frames, outputting a high-dimensional supplementary visual feature tensor. Simultaneously, the cross-view tensor fusion recalculation unit extracts the initial visual feature tensor of the main control PTZ camera's image at the subgraph path association time. To ensure the mathematical consistency of the tensor structure, tensor pruning and scaling resampling are performed to ensure absolute alignment of the supplementary visual feature tensor with the initial visual feature tensor in the length, width, and depth dimensions. After alignment, the cross-view tensor fusion recalculation unit performs channel alignment concatenation along the feature channel depth axis to generate a fused feature map containing dual-view spatial information. Subsequently, this fused feature map is input into a fully connected classification network for probability matrix calculation. A multidimensional logistic regression activation function is used to process the output value of this probability matrix to calculate the joint probability distribution. Each element of this joint probability distribution represents a normalized probability value of the corresponding violation type. Finally, the violation category corresponding to the element with the largest value in the joint probability distribution, along with the probability value itself, is extracted to generate a supplementary evidence assessment result containing a high-confidence conclusion. For example... Figure 4 As shown, the data span offset of the dumbbell diagram can verify that after introducing the supplementary visual features of the adjacent collaborative PTZ camera to perform tensor channel alignment stitching, the confidence probability of the target violation category is significantly corrected and absolutely increased, eliminating the potential for misjudgment caused by physical obstruction from a single monitoring perspective.

[0037] Optionally, the system further includes: Based on the frequency of violations determined by the initial violation assessment, the spatial range of the work area boundary coordinates is expanded, and a dynamic warning and defense line level is generated for the corresponding construction area.

[0038] The system retrieves the initial violation judgment results recorded in the system database. It then extracts the violation frequency parameter, which represents the cumulative total of all confirmed violation events occurring within the specified time period in the work area. Since high-frequency violation areas often indicate structural weaknesses in on-site safety management, this module triggers a defensive spatial protection mechanism, expanding the spatial range of the work area boundary coordinates. An expansion factor is introduced; this factor is not subjectively set but derived from the geometric series tolerance derived from the violation scoring increment rule. The calculation relationship for the spatial range expansion scalar satisfies the following formula: , Among them, letters Represents the calculated spatial extent extended by a scalar, in meters; letters Represents the frequency of traffic violations extracted from the database; letters This represents the outward expansion ratio coefficient. Subsequently, the outward expansion scalar of this spatial range is superimposed onto the outward normal direction of each latitude and longitude boundary segment of the original work area boundary coordinates, generating a completely new two-dimensional safety fence after outward translation. Simultaneously, based on the absolute value range of the outward expansion scalar, a dynamic alert defense level identified by a numerical sequence is generated. For example, an outward expansion scalar below a certain threshold is classified as a light defense level, and one above the threshold as a heavy defense level. This defense level parameter is forcibly issued to subsequent operational flows to guide the positioning distance and patrol density of on-site monitoring personnel.

[0039] The violation alarm dispatch module is used to map the violation type tag of the supplementary evidence collection and judgment result to the violation knowledge base data to extract the rectification time limit parameter, extract the violation video frame segment of the supplementary evidence collection and judgment result, encapsulate the violation video frame segment, the rectification time limit parameter and the boundary coordinate of the work area, and generate violation alarm dispatch data. Optionally, the traffic violation alarm dispatch module includes: The time interval interception unit is used to parse the violation type label of the supplementary evidence collection judgment result, perform entity index matching between the violation type label and the violation knowledge base data, extract the rectification time limit parameter, parse the start timestamp and end timestamp of the violation type label, and intercept the continuous video image from the start timestamp to the end timestamp in the supplementary evidence collection judgment result to generate violation video frame segments. The temporal interval extraction unit parses the violation type label representing the specific violation from the supplementary evidence judgment results output by the cross-view tensor fusion recalculation unit. To obtain the standard handling requirements for this type of violation, the temporal interval extraction unit performs a matching operation between the violation type label and the primary key entity index of the violation knowledge base data through the underlying database query interface. This violation knowledge base data is a set of rule mappings constructed using relational data tables. The rectification time limit parameters stored internally are not arbitrarily set manually, but are strictly considered based on the graded assessment of the danger level of different violations, and are mapped to fixed standard hourly values. For example, serious violations endangering personal safety will be objectively mapped to extremely short rectification time limit parameters, while violations related to civilized construction will be mapped to longer rectification time limit parameters. After matching, the temporal interval extraction unit accurately extracts the rectification time limit parameter. Simultaneously, to solidify the complete visual evidence chain of the violation, the temporal interval extraction unit deeply analyzes the original temporal sequence bound to the violation type label, extracting the start timestamp corresponding to the first frame of the violation action and the end timestamp corresponding to the last frame of the violation action. Considering that surveillance videos may exhibit pre- and post-motion transitions when recording continuous human movements, simply extracting footage based on start and end timestamps could lead to abrupt changes in the evidence or loss of crucial contextual information. Therefore, the time-series interval extraction unit introduces a time-series buffer constant, objectively fixed at 3000 milliseconds based on the average physiological reaction time of human visual short-term memory. The time-series interval extraction unit shifts the start timestamp forward by 3000 milliseconds and the end timestamp backward by 3000 milliseconds, using the extended time interval as the physical addressing range. This allows for precise extraction of a complete, continuous video segment from the original surveillance storage stream, generating a violation video frame segment containing the complete context of the violation.

[0040] The dynamic routing encapsulation unit is used to extract the boundary coordinates of the work area and the dynamic warning defense level, convert the dynamic warning defense level into a dispatch route weight, and perform markup language serialization encoding on the violation video frame, the rectification time limit parameter and the boundary coordinates of the work area, and fuse the dispatch route weight to generate violation alarm dispatch data.

[0041] The dynamic routing encapsulation unit acquires the boundary coordinates of the work area and the dynamic alert level. The dynamic alert level is a scalar of hazard severity, identified by integers. To ensure that alarms from high-risk sites reach security management personnel's terminals first, especially under extreme conditions such as limited network bandwidth or a surge in alarm concurrency, the dynamic routing encapsulation unit converts this dynamic alert level into route weights. The conversion calculation relationship for route weights satisfies the following mathematical formula: , Among them, letters This represents the calculated dispatch route weight, a dimensionless floating-point number that directly determines the queuing priority of alarm data packets in the message middleware queue; the letter... Integer scalars representing the levels of dynamic alert defense; letters This represents the emergency allocation coefficient, whose value is derived from the ratio of the maximum allowed number of high-speed network threads to the total physical quota of concurrent threads in the backend dispatch server; the letter... This represents a basic guarantee constant, used to ensure that even the lowest-level alarms are allocated the minimum channel weight required to maintain transmission. Its value is calculated based on the minimum transmission throughput of the server's network interface card. After calculating the dispatch route weights, the dynamic routing encapsulation unit performs markup language serialization encoding on the acquired multi-source heterogeneous data. This encoding process converts violation video frames into plain text binary strings using a Base64 encoding algorithm to avoid transmission truncation, and converts the rectification time limit parameters and work area boundary coordinates into key-value pair nodes in JSON or XML standard format. Finally, the dynamic routing encapsulation unit merges the calculated dispatch route weights and writes them into the header control field of the serialized message, thereby generating complete, priority-carrying, and network-wide universal violation alarm dispatch data.

[0042] For example, the dynamic routing encapsulation unit receives an alarm command for an out-of-bounds event. It extracts the dynamic alert level generated by the front-end module and parses its integer scalar value. The value is 1. Simultaneously, the hardware resource monitoring interface of the backend dispatch server is retrieved, and the current high-speed network thread physical quota ratio is read as 0.6. Therefore, the emergency allocation coefficient is set to 1. The value is set to 0.6; the basic guarantee constant is calculated from the minimum transmission throughput of the network card. The value is 0.2. Substituting the above value into the dispatch route weight calculation formula, the numerator 0.6 divided by the denominator 1 yields 0.6, which is then added to the basic guarantee constant 0.2 to finally calculate the dispatch route weight. The weight is 0.8. During the data encoding phase, the dynamic routing encapsulation unit converts the 9-second-long violation video frame segment into a very long character stream composed of letters and numbers using the Base64 algorithm; formats the rectification time limit parameter of 0.5 hours into a floating-point numerical attribute; and formats the work area boundary coordinates into a comma-separated latitude and longitude array attribute. Subsequently, markup language serialization encoding is performed to construct a standard JSON text block. In the protocol header "Header" area of ​​this JSON text block, a weight control tag with the attribute name "RouteWeight" and an attribute value of 0.8 is written. After the generated violation alarm dispatch data is pushed into the network switch, the network router reads the dispatch routing weight as high as 0.8 and immediately opens a high-priority service quality dedicated transmission channel for it, ensuring that the dispatch data is pushed to the project chief engineer's handheld mobile terminal with extremely low latency at the millisecond level.

[0043] The rectification review and archiving module is used to receive rectification feedback instructions for the violation alarm dispatch data, drive the associated PTZ camera of the work area boundary coordinates to acquire the review video stream, calculate the defect elimination fitting degree between the visual features of the review video stream and the supplementary evidence collection judgment result, and output closed-loop archived logs based on the defect elimination fitting degree.

[0044] Optionally, the rectification review and archiving module includes: The verification drive and field of view calibration unit is used to parse the trigger timestamp of the rectification feedback instruction, map the boundary coordinates of the work area to the PTZ directional control matrix, and drive the associated PTZ camera to obtain the verification video stream based on the PTZ directional control matrix at the trigger timestamp. The verification driver and field-of-view calibration unit establish a low-level socket listening service to receive rectification feedback instructions for violation alarm dispatch data issued by the mobile inspection terminal or the on-site management platform. This rectification feedback instruction is a data packet containing the processing status. The verification driver and field-of-view calibration unit parse the protocol header of this data packet to accurately extract the trigger timestamp indicating that the rectification action has been confirmed as completed. This trigger timestamp records the absolute moment of manual feedback processing. Simultaneously, based on the traceability serial number carried in the data packet, the verification driver and field-of-view calibration unit extract the corresponding work area boundary coordinates from the violation alarm dispatch data cached in the background. To achieve "unmanned" automatic acceptance of rectification results by machine vision, the verification driver and field-of-view calibration unit maps the work area boundary coordinates on this two-dimensional plane to a pan-tilt-zoom (PTZ) directional control matrix, directing the hardware equipment in the form of commands. This mapping process first calls the on-site infrastructure network topology table to query the network PTZ camera that is spatially closest to the work area boundary coordinates and has no building obstruction in its line of sight as the associated PTZ camera. Subsequently, the verification drive and field-of-view calibration unit extracts the extreme values ​​of latitude and longitude at the two ends of the diagonal of the boundary coordinates of the work area, and calculates the physical centroid plane coordinates of the work area using an arithmetic mean. The verification drive and field-of-view calibration unit extracts the three-dimensional world coordinates of the associated PTZ camera due to its fixed base, and calculates the horizontal deflection angle and vertical pitch angle parameters of the aforementioned physical centroid plane coordinates relative to these three-dimensional world coordinates. These two angle parameters are derived from trigonometric functions of a right triangle constructed by calculating the difference between the ground projection distance and the tower elevation. The verification drive and field-of-view calibration unit encapsulates the calculated horizontal deflection angle and vertical pitch angle parameters according to the underlying protocol standard format to generate a PTZ orientation control matrix. Finally, the verification drive and field-of-view calibration unit sends this PTZ orientation control matrix to the associated PTZ camera via the network, forcibly driving the servo motor of the associated PTZ camera to adjust the lens attitude at the triggered timestamp, aiming at the rectified site area to capture continuous image data under this time slice, thus generating the verification video stream.

[0045] The inter-period feature comparison and calculation unit is used to extract the current visual feature vector of the reviewed video stream, extract the benchmark violation feature vector of the supplementary evidence judgment result, calculate the spatial difference scalar between the current visual feature vector and the benchmark violation feature vector, generate the defect elimination fitting degree, encapsulate it into a structured data block, and output closed-loop archived log.

[0046] The inter-period feature comparison and resolution unit receives the review video stream from the review driver and field-of-view calibration unit. To quantify the difference in on-site conditions before and after rectification, the inter-period feature comparison and resolution unit extracts a key frame from the review video stream, inputs it into an image convolutional neural network, extracts its deep semantic features, and performs global average pooling to extract a high-dimensional floating-point array representing the current on-site state, i.e., the current visual feature vector. Simultaneously, the inter-period feature comparison and resolution unit retrieves the supplementary evidence collection and judgment results for the violation event by calling the memory index, extracting the original on-site feature array saved when the violation was determined, i.e., the baseline violation feature vector. Both feature vectors are generated under the same network architecture and have completely identical dimensional lengths, ensuring the legality of the mathematical operations. The inter-period feature comparison and resolution unit calculates the spatial difference scalar between the current visual feature vector and the baseline violation feature vector. The calculation relationship of this spatial difference scalar adopts the Euclidean distance metric model in high-dimensional vector space, which satisfies the following mathematical formula: , Among them, letters This represents the calculated spatial difference scalar, which is a non-negative floating-point number; the letter... The numerical value representing the uniform channel depth dimension of the feature vector; letters Represents the current visual feature vector at the th Component values ​​on each feature channel; letters The representative baseline violation feature vector is in the th The component values ​​on each feature channel. If the on-site hazard is completely corrected, the deep features of the reviewed image will significantly change compared to the deep features of the violation image, and the sum of squares of the differences between the components in each channel will be drastically amplified, leading to... The output is a relatively large value. After obtaining the spatial difference scalar, the inter-period feature comparison and solution unit uses an sigmoid nonlinear activation function to normalize the spatial difference scalar to a probability range of 0 to 1, generating a defect elimination fit degree. The closer the defect elimination fit degree is to 1, the greater the change in the on-site state and the more thorough the rectification. Finally, the inter-period feature comparison and solution unit performs key-value pair encoding on the calculated defect elimination fit degree, trigger timestamp, work area boundary coordinates, and review video stream segments, encapsulating them into a structured data block. This structured data block is written to the audit database of the backend safety management system in an immutable format, thereby generating a complete, compliant, and closed-loop archived log with objective machine vision acceptance conclusions.

[0047] Based on the same inventive concept, this invention also provides a method for intelligent identification and closed-loop handling of safety violations at power transmission and transformation engineering construction sites, the method comprising: Acquire PTZ camera video stream, engineering tower coordinate data, work plan data, and violation knowledge base data; perform frame extraction operation on the PTZ camera video stream to extract video frame sequences; parse the engineering tower coordinate data and the work plan data to extract the work area boundary coordinates and work time window; and construct a construction scene dataset. The video frame sequence and work area boundary coordinates are extracted using the construction scene dataset. Multi-target tracking is performed on the video frame sequence. The cross-frame dynamic distance between the personnel node displacement vector and the work area boundary coordinates is calculated. A temporal behavior directed graph is constructed using the cross-frame dynamic distance as the edge attribute. Extract the safety regulation condition topology from the violation knowledge base data, combine it with the operation time window, perform distance calculation on the directed graph of the time sequence behavior, extract the subgraph path that deviates from the safety regulation condition topology, and output the initial violation judgment result; Based on the initial violation determination result, the three-dimensional spatial location is parsed, a field of view adjustment command is generated and sent to the adjacent cooperative PTZ camera at the three-dimensional spatial location, and the joint probability distribution is calculated by fusing the supplementary video frames returned by the adjacent cooperative PTZ camera to generate a supplementary evidence determination result. Based on the violation type tag of the supplementary evidence collection and judgment result, the rectification time limit parameter is extracted from the violation knowledge base data, the violation video frame segment of the supplementary evidence collection and judgment result is extracted, the violation video frame segment, the rectification time limit parameter and the boundary coordinate of the work area are encapsulated, and violation alarm dispatch data is generated. The system receives rectification feedback instructions for the violation alarm dispatch data, drives the associated PTZ camera at the boundary coordinates of the work area to acquire the review video stream, calculates the defect elimination fitting degree between the visual features of the review video stream and the supplementary evidence collection judgment result, and outputs closed-loop archived logs based on the defect elimination fitting degree.

[0048] It should be noted that the formulas described above, through the principle of dimensional consistency and mathematical standardization methods (such as normalization, dimensionless parameter conversion, or unit system unification), can translate physical quantities with different properties into unitless standard values ​​or superimposed parameters of the same dimension. This eliminates the interference of different dimensions on the computational logic, allowing the formulas to retain the original data distribution characteristics while possessing mathematical rationality and adaptability to objective laws. These are conventional technical methods and will not be elaborated upon here. The electrical connections between the various units described above do not necessarily represent direct or indirect connections; any indirect connection method is applicable to the embodiments of this invention as long as it achieves the purpose of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of this invention.

[0049] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A smart identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites, characterized in that: The system includes: The construction scenario dataset construction module is used to acquire PTZ camera video stream, engineering tower coordinate data, work plan data and violation knowledge base data, perform frame extraction operation on the PTZ camera video stream to extract video frame sequences, and parse the engineering tower coordinate data and the work plan data to extract the work area boundary coordinates and work time window to construct the construction scenario dataset. The temporal behavior directed graph construction module is used to extract video frame sequences and work area boundary coordinates using the construction scene dataset, perform multi-target tracking on the video frame sequences, calculate the cross-frame dynamic distance between the personnel node displacement vector and the work area boundary coordinates, and construct a temporal behavior directed graph with the cross-frame dynamic distance as the edge attribute. The initial violation determination module is used to extract the safety regulation condition topology of the violation knowledge base data, combine it with the operation time window, perform distance calculation on the directed graph of the time sequence behavior, extract the subgraph path that deviates from the safety regulation condition topology, and output the initial violation determination result. The PTZ camera supplementary evidence collection module is used to parse the three-dimensional spatial position based on the initial violation judgment result, generate a field of view adjustment command and send it to the adjacent cooperative PTZ cameras at the three-dimensional spatial position, fuse the return supplementary video frames of the adjacent cooperative PTZ cameras to calculate the joint probability distribution, and generate a supplementary evidence collection judgment result. The violation alarm dispatch module is used to map the violation type tag of the supplementary evidence collection and judgment result to the violation knowledge base data to extract the rectification time limit parameter, extract the violation video frame segment of the supplementary evidence collection and judgment result, encapsulate the violation video frame segment, the rectification time limit parameter and the boundary coordinate of the work area, and generate violation alarm dispatch data. The rectification review and archiving module is used to receive rectification feedback instructions for the violation alarm dispatch data, drive the associated PTZ camera of the work area boundary coordinates to acquire the review video stream, calculate the defect elimination fitting degree between the visual features of the review video stream and the supplementary evidence collection judgment result, and output closed-loop archived logs based on the defect elimination fitting degree.

2. The intelligent identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites according to claim 1, characterized in that, The construction scenario dataset construction module includes: The video feature processing unit is used to parse the temporal coding structure of the PTZ camera video stream, perform dynamic keyframe extraction, strip static background images, and generate a video frame sequence. The multi-dimensional spatiotemporal fusion unit is used to extract the three-dimensional geodetic coordinate values ​​of the engineering tower coordinate data, perform polygon bounding box spatial mapping in combination with the scheduling field in the work plan data, generate the boundary coordinates of the work area and the work time window, and stitch the video frame sequence to construct the construction scene dataset.

3. The intelligent identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites according to claim 1, characterized in that, The directed graph construction module for temporal behavior includes: The entity trajectory calculation unit is used to extract the video frame sequence of the construction scene dataset, perform pixel-level instance segmentation and cross-frame feature point matching on the video frame sequence, extract the centroid coordinates of the target bounding boxes of consecutive frames, and calculate the temporal difference of the centroid coordinates to generate personnel node displacement vectors. The topology graph construction unit is used to extract the boundary coordinates of the work area in the construction scene dataset, calculate the shortest spatial distance between the personnel node displacement vector and the boundary coordinates of the work area, generate cross-frame dynamic distance, extract the personnel node displacement vector as graph node attribute and map the cross-frame dynamic distance as edge connection weight, and construct a temporal behavior directed graph.

4. The intelligent identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites according to claim 1, characterized in that, The initial violation determination module includes: The spatiotemporal graph mapping unit is used to extract the safety regulation condition topology of the violation knowledge base data and the operation time window of the construction scenario dataset, perform the node timestamp embedding mapping from the operation time window to the safety regulation condition topology, and construct the benchmark constraint graph. The graph editing distance calculation unit is used to perform graph editing distance calculation on the temporal behavior directed graph and the benchmark constraint graph, filter abnormal node sequences, extract the connected topology of the abnormal node sequences as subgraph paths that deviate from the topology of the safety regulations, and output the initial violation judgment result.

5. The intelligent identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites according to claim 1, characterized in that, The PTZ camera supplementary evidence collection module includes: The spatial geometric perspective positioning unit is used to extract the two-dimensional pixel coordinate matrix of the initial violation judgment result, perform the perspective projection transformation of the two-dimensional pixel coordinate matrix to the world coordinate system, resolve the three-dimensional spatial position, calculate the gimbal rotation vector and zoom factor of the three-dimensional spatial position and the adjacent cooperative PTZ camera, generate the field of view adjustment command and send it out. The cross-viewpoint tensor fusion recalculation unit is used to receive the supplementary video frames returned by the adjacent cooperative PTZ camera based on the field of view adjustment command, extract the supplementary visual feature tensors of the returned supplementary video frames, align and splice the supplementary visual feature tensors with the initial visual feature execution channel of the subgraph path, calculate the joint probability distribution, and generate supplementary evidence collection judgment results.

6. The intelligent identification and closed-loop handling system for safety violations at power transmission and transformation construction sites according to claim 1, characterized in that, The system also includes: Based on the frequency of violations determined by the initial violation assessment, the spatial range of the work area boundary coordinates is expanded, and a dynamic warning and defense line level is generated for the corresponding construction area.

7. The intelligent identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites according to claim 6, characterized in that, The violation alarm dispatch module includes: The time interval interception unit is used to parse the violation type label of the supplementary evidence collection judgment result, perform entity index matching between the violation type label and the violation knowledge base data, extract the rectification time limit parameter, parse the start timestamp and end timestamp of the violation type label, and intercept the continuous video image from the start timestamp to the end timestamp in the supplementary evidence collection judgment result to generate violation video frame segments. The dynamic routing encapsulation unit is used to extract the boundary coordinates of the work area and the dynamic warning defense level, convert the dynamic warning defense level into a dispatch route weight, and perform markup language serialization encoding on the violation video frame, the rectification time limit parameter and the boundary coordinates of the work area, and fuse the dispatch route weight to generate violation alarm dispatch data.

8. The intelligent identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites according to claim 1, characterized in that, The rectification review and archiving module includes: The verification drive and field of view calibration unit is used to parse the trigger timestamp of the rectification feedback instruction, map the boundary coordinates of the work area to the PTZ directional control matrix, and drive the associated PTZ camera to obtain the verification video stream based on the PTZ directional control matrix at the trigger timestamp. The inter-period feature comparison and calculation unit is used to extract the current visual feature vector of the reviewed video stream, extract the benchmark violation feature vector of the supplementary evidence judgment result, calculate the spatial difference scalar between the current visual feature vector and the benchmark violation feature vector, generate the defect elimination fitting degree, encapsulate it into a structured data block, and output closed-loop archived log.

9. A method for intelligent identification and closed-loop handling of safety violations at power transmission and transformation engineering construction sites, applied to the intelligent identification and closed-loop handling system for safety violations at power transmission and transformation engineering construction sites as described in any one of claims 1-8, characterized in that, The method includes: Acquire PTZ camera video stream, engineering tower coordinate data, work plan data, and violation knowledge base data; perform frame extraction operation on the PTZ camera video stream to extract video frame sequences; parse the engineering tower coordinate data and the work plan data to extract the work area boundary coordinates and work time window; and construct a construction scene dataset. The video frame sequence and work area boundary coordinates are extracted using the construction scene dataset. Multi-target tracking is performed on the video frame sequence. The cross-frame dynamic distance between the personnel node displacement vector and the work area boundary coordinates is calculated. A temporal behavior directed graph is constructed using the cross-frame dynamic distance as the edge attribute. Extract the safety regulation condition topology from the violation knowledge base data, combine it with the operation time window, perform distance calculation on the directed graph of the time sequence behavior, extract the subgraph path that deviates from the safety regulation condition topology, and output the initial violation judgment result; Based on the initial violation determination result, the three-dimensional spatial location is parsed, a field of view adjustment command is generated and sent to the adjacent cooperative PTZ camera at the three-dimensional spatial location, and the joint probability distribution is calculated by fusing the supplementary video frames returned by the adjacent cooperative PTZ camera to generate a supplementary evidence determination result. Based on the violation type tag of the supplementary evidence collection and judgment result, the rectification time limit parameter is extracted from the violation knowledge base data, the violation video frame segment of the supplementary evidence collection and judgment result is extracted, the violation video frame segment, the rectification time limit parameter and the boundary coordinate of the work area are encapsulated, and violation alarm dispatch data is generated. The system receives rectification feedback instructions for the violation alarm dispatch data, drives the associated PTZ camera at the boundary coordinates of the work area to acquire the review video stream, calculates the defect elimination fitting degree between the visual features of the review video stream and the supplementary evidence collection judgment result, and outputs closed-loop archived logs based on the defect elimination fitting degree.

Images