Visual early warning system for preventing personnel from entering into power transmission line construction by mistake

Abstract

The invention relates to the technical field of power transmission line construction safety monitoring and computer vision digital twinning, in particular to a power transmission line construction personnel mistaken entry prevention visual early warning system which comprises a data acquisition and reference reconstruction step: generating a reference three-dimensional projection image and a reference depth map which are consistent in camera visual angle and free of personnel interference; a real difference extraction step: generating a real-time monitoring difference feature map containing unexpected changes; a theoretical simulation deduction step: generating a simulation intrusion feature map representing a human body intrusion form; a feature coupling judgment step: when the spatial coincidence degree is greater than a preset spatial consistency threshold value, judging that a real person mistaken entering behavior exists and triggering an alarm signal; if the spatial coincidence degree is smaller than or equal to a spatial consistency threshold value, determining that the noise is environmental noise and keeping a current monitoring state; the technical problem that traditional static background modeling cannot adapt to frequent change of a construction site scene is solved, false alarm caused by normal construction activities is avoided, and monitoring pertinence is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring and computer vision digital twin technology for power transmission line construction, specifically a visual early warning system for preventing personnel from accidentally entering power transmission line construction sites. Background Technology

[0002] With the continuous expansion of power grid construction, the environment at transmission line construction sites is becoming increasingly complex; this complexity poses a severe challenge to construction safety management, especially to monitoring and preventing personnel from accidentally entering dangerous areas. Currently, security monitoring at construction sites generally employs manual polling of video surveillance or traditional image recognition technology. Monitoring personnel identify potential safety hazards by observing the surveillance footage, or the system uses algorithms such as static background modeling to detect the presence of moving targets within the monitoring range. Transmission line construction sites are highly dynamic environments, involving frequent movement of construction machinery, changes in the location of tower materials, and the influence of weather conditions such as lighting and fog. Traditional background modeling methods are ill-suited to these legitimate scene changes, easily misinterpreting minor disturbances, changes in light and shadow, or planned machinery operations as intrusions. Furthermore, relying solely on manual monitoring cannot guarantee 24 / 7 focused attention, resulting in high false alarm rates and delayed responses, making it difficult to accurately identify genuine personnel intrusions in complex and interfering environments. Therefore, effectively eliminating environmental interference in dynamically changing construction environments and achieving accurate, real-time early warnings of unauthorized personnel intrusions has become an urgent problem to be solved in this field. Summary of the Invention

[0003] To solve the above-mentioned technical problems, the present invention provides a visual early warning system for preventing personnel from accidentally entering power transmission line construction sites. Specifically, the technical solution of the present invention includes: The processor, memory, and communication interface; the memory stores a computer program, and the processor executes the computer program to perform the following steps: Data acquisition and benchmark reconstruction steps: acquire BIM model data of transmission line construction, GIS geographic information, construction schedule and camera internal and external parameters, calculate solar altitude angle based on current time, and use rasterization rendering technology to render and generate a benchmark 3D projection image and benchmark depth map that are consistent with the camera's viewpoint and free from human interference. Reality difference extraction steps: Real-time monitoring video frames of the construction site are acquired by the camera, the pixel differences and depth differences between the real-time monitoring video frames and the reference three-dimensional projection image are calculated, and a real-time monitoring differential feature map containing unexpected changes is generated. Theoretical simulation deduction steps: Based on the human skeletal kinematic parameters in the preset risk parameter library, a virtual three-dimensional human model is generated in the three-dimensional space corresponding to the benchmark three-dimensional projection image, and a simulated invasion feature map representing the human invasion pattern is generated through physical rendering. Feature coupling decision steps: Calculate the spatial overlap between the real-time monitoring differential feature map and the simulated intrusion feature map in the same pixel coordinate system; if the spatial overlap is greater than the preset spatial consistency threshold, it is determined that there is real personnel trespassing and an alarm signal is triggered; if the spatial overlap is less than or equal to the spatial consistency threshold, it is determined to be environmental noise and the current monitoring state is maintained.

[0004] Preferably, rasterization rendering technology is used to generate a reference 3D projection image and a reference depth map that are consistent with the camera's viewpoint and free from human interference. Specifically, this includes: parsing the legal equipment location coordinates and tower material stacking point coordinates of the current construction stage according to the construction schedule, and updating the BIM model layout; combining the GIS geographic information and the current time point, calculating the solar azimuth and elevation angles, and generating a scene shadow mask; constructing a virtual camera view frustum using the camera's intrinsic and extrinsic parameters, and combining the scene shadow mask, ambient light color extracted from the sky area of ​​the real-time monitoring video frame, and current meteorological visibility parameters to perform rasterization rendering on the updated BIM model, outputting the reference 3D projection image containing RGB color channels, and the reference depth map containing distance information.

[0005] Preferably, the calculation of pixel differences and depth differences between the real-time monitoring video frame and the reference 3D projection image to generate a real-time monitoring differential feature map containing unexpected changes specifically includes: performing semantic segmentation on the real-time monitoring video frame and the reference 3D projection image respectively to obtain semantic label maps; performing consistency comparison operations on the semantic label maps to identify difference regions with inconsistent semantic categories, and based on the construction plan table, removing pixel regions belonging to legitimate construction machinery movement from the difference regions; extracting pixel-level residual values ​​and depth-level residual values ​​from the remaining difference regions, and generating the real-time monitoring differential feature map by stitching them together.

[0006] Preferably, generating a virtual three-dimensional human body model in the three-dimensional space corresponding to the reference three-dimensional projection image specifically includes: reading the safety distance value in the power safety work regulations, delineating the prohibited three-dimensional space area in the BIM model; generating a set of virtual three-dimensional human body models based on human skeletal kinematic constraints, placing the virtual three-dimensional human body models on the ground grid in the prohibited three-dimensional space area, and simulating their occlusion relationship and motion posture under the current camera viewpoint according to the camera perspective projection relationship.

[0007] Preferably, generating a simulated intrusion feature map representing the morphology of human intrusion through physical rendering specifically includes: rendering a scene with a virtual three-dimensional human model as a simulated disturbed image, and calculating the pixel difference between the simulated disturbed image and the reference three-dimensional projection image; extracting the texture features and contour features of the pixel change area caused only by the virtual three-dimensional human model, and generating the simulated intrusion feature map.

[0008] Preferably, calculating the spatial overlap between the real-time monitoring differential feature map and the simulated intrusion feature map in the same pixel coordinate system specifically includes: extracting the edge gradient magnitude features of the corresponding region of the real-time monitoring differential feature map using the Sobel operator; extracting the geometric contour features of the virtual three-dimensional human model in the simulated intrusion feature map; calculating the intersection-over-union (IoU) value of the edge gradient magnitude features and the geometric contour features in the pixel coordinate system, and using the IoU value as the spatial overlap.

[0009] Preferably, the processor is further configured to perform a temporal causality verification step: extracting optical flow field data from multiple consecutive frames of the real-time monitoring video frames, and calculating displacement vectors in the difference regions; obtaining the simulated motion trajectory vector of the virtual three-dimensional human model within a preset time window, and calculating the cosine similarity between the displacement vector and the simulated motion trajectory vector; if the cosine similarity is greater than a preset motion law threshold, then the temporal causality verification is confirmed to have passed; if the cosine similarity is less than or equal to the motion law threshold, then the temporal causality verification is confirmed to have failed.

[0010] Preferably, determining the existence of a genuine unauthorized entry and triggering an alarm signal specifically includes: generating an alarm signal containing the coordinates and time of the intrusion location only when the spatial overlap is greater than the spatial consistency threshold and the temporal causality check passes, and sending it to the on-site audible and visual alarm; if the spatial overlap is less than or equal to the spatial consistency threshold, or the temporal causality check fails, then determining that the real-time monitoring differential feature map is environmental background noise, not generating an alarm signal, and recording the current log as a false alarm event.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves automatic adaptation to legitimate changes at the construction site by introducing a dynamic update mechanism for the building information model driven by the construction schedule. The system, combined with semantic comparison operations, can accurately identify and eliminate legitimate dynamic interferences within the plan, such as mechanical operations and material stacking. This effectively solves the technical problem that traditional static background modeling cannot adapt to frequent changes in the construction site scene, avoids false alarms caused by normal construction activities, and ensures the targeting of monitoring. 2. This invention utilizes physically based rendering technology, combined with geographic information and real-time meteorological parameters, to construct a benchmark holographic field that is strictly consistent with the actual ambient lighting and visibility. By simulating solar shadows and atmospheric transmittance, the system can eliminate the nonlinear interference of drastic changes in light and shadow and hazy weather on image contrast. This significantly improves the robustness of the system under complex outdoor meteorological conditions and solves the problem that optical monitoring is easily affected by environmental noise. 3. This invention employs a theoretical simulation and feature-coupled decision method based on skeletal kinematics, achieving a leap from passive identification to active verification. By calculating the spatial overlap between real differential features and virtual 3D human models, the system can physically distinguish between real people and non-human forms such as swaying trees. This morphological fingerprint alignment mechanism greatly reduces the interference of false targets and improves the accuracy of early warning. 4. This invention constructs a dual-logic closed loop that includes temporal causality verification. By comparing the similarity between the real-time optical flow field and the simulated motion trajectory, the motion law of the intrusion behavior is verified. An alarm is triggered only when the target conforms to human physical characteristics in both spatial shape and motion trend, effectively suppressing misjudgments caused by instantaneous similarity or disordered motion. This strict logic and operation mechanism ensures the high reliability of the alarm signal and reduces unnecessary interference with the construction progress. Attached Figure Description

[0012] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] Example 1: Please see Figure 1 A visual early warning system for preventing unauthorized entry during power transmission line construction includes: a processor, a memory, and a communication interface; the memory stores a computer program, and the processor executes the computer program to perform the following steps: Data acquisition and benchmark reconstruction steps: acquire BIM model data of transmission line construction, GIS geographic information, construction schedule and camera internal and external parameters, calculate solar altitude angle based on current time, and use physically based rasterization rendering technology to generate benchmark 3D projection image and benchmark depth map that are consistent with the camera's viewpoint and free from human interference. Reality difference extraction steps: Real-time monitoring video frames of the construction site are captured by cameras, and the pixel differences and depth differences between the real-time monitoring video frames and the reference 3D projection image are calculated to generate a real-time monitoring differential feature map containing unexpected changes. Theoretical simulation deduction steps: Based on the human skeletal kinematic parameters in the preset risk parameter library, a virtual three-dimensional human model is generated in the three-dimensional space corresponding to the benchmark three-dimensional projection image, and a simulated invasion feature map representing the human invasion pattern is generated through physical rendering. Feature coupling decision steps: Calculate the spatial overlap between the real-time monitoring differential feature map and the simulated intrusion feature map in the same pixel coordinate system; if the spatial overlap is greater than the preset spatial consistency threshold, it is determined that there is real personnel trespassing and an alarm signal is triggered; if the spatial overlap is less than or equal to the spatial consistency threshold, it is determined to be environmental noise and the current monitoring status is maintained.

[0015] This embodiment provides a visual early warning system for preventing personnel from accidentally entering power transmission line construction sites. Its hardware architecture is based on edge computing or cloud server architecture, mainly including a processor, such as a high-performance GPU / TPU cluster, a memory for storing BIM model database, GIS database and algorithm model, and a communication interface for connecting on-site monitoring cameras and alarm devices. The system's workflow is not traditional image recognition, but rather based on a closed-loop logic of forward and reverse dual verification. Specifically, the processor executes the following core steps: Data Acquisition and Baseline Reconstruction: This step aims to generate a pure, ideal holographic field. The baseline 3D projection image refers to a 2D image reconstructed in virtual space using computer graphics technology, which eliminates all unintended interference, such as unauthorized personnel and unplanned vehicles, but retains all legitimate environmental features, such as terrain, planned equipment placement, and shadows cast under the current lighting conditions. Implementation logic: The processor acquires BIM model data for the power transmission line construction via a communication interface, including geometric information of the towers, foundations, and fences, GIS geographic information (topography and elevation data), and a construction schedule. This construction schedule is a key constraint input, defining the legal construction phases and machinery positions at the current time. Simultaneously, it acquires the camera's intrinsic focal length, principal point, and extrinsic rotation matrix. Translation vector Based on the current system time, the solar altitude angle and azimuth angle are calculated using astronomical algorithms; physical rasterization rendering technology is used, combined with the principles of shadow mapping and perspective projection, to render an image in virtual space that is completely consistent with the perspective of the on-site camera; the output generates a reference 3D projection image RGB channel and reference depth map Z-buffer information. Reality Difference Extraction: This step aims to extract suspicious signals from the chaotic reality; Real-time monitoring differential feature map refers to a high-dimensional dataset that represents the difference between the real scene and the ideal benchmark. It removes the static background and retains only the unexpected changes. Implementation logic: Cameras capture real-time monitoring video frames from the construction site; the processor calculates the pixel-level RGB value differences and geometric differences between this frame and the reference 3D projection image, i.e., the numerical differences between the real-time depth map estimated based on deep learning and the reference depth map; this process is essentially a data cleaning process, stripping away the complex construction site background; Mathematical model: Let the real-time frame be... The baseline diagram is Then the difference features ,in, This is a preprocessing extraction function that includes semantic segmentation validity mask filtering, 5×5 Gaussian filtering smoothing, and 3×3 morphological opening operation denoising. Theoretical simulation and deduction: This step aims to actively generate the expected intrusion fingerprint based on physical rules; the simulated intrusion feature map refers to the visual features that the intrusion behavior should present on the image, which the system deduces based on human biomechanics and perspective projection principles when an intrusion behavior is assumed to exist. Implementation logic: Unlike passive recognition, this system uses a preset risk parameter library, which includes skeletal kinematic parameters such as human height, stride, and joint degrees of freedom, to actively generate a virtual 3D human model in the virtual 3D space corresponding to the baseline 3D projection image; through physical rendering (PBR), it simulates the form of the virtual human under the current lighting and occlusion conditions and generates feature vectors that characterize the human invasion form. Feature Coupling Decision: This step is the alignment verification of virtual and real fingerprints; spatial overlap refers to the degree of similarity in topological structure between the real-time monitoring differential feature map of abnormal changes observed in reality and the simulated intrusion feature map of the theoretically deduced intrusion pattern; implementation logic: the processor calculates the degree of overlap between the above two vectors in the same pixel coordinate system; if the spatial overlap is greater than the spatial consistency threshold... This indicates that the abnormal changes in reality conform to human physical characteristics, and are determined to be the result of a real person mistakenly entering the space, triggering an alarm; if the spatial overlap is less than... This indicates that abnormal changes in reality, such as swaying trees or fluttering plastic bags, cannot be explained by the human body model and are identified as environmental noise, which should be monitored. This embodiment solves the problem that traditional background modeling cannot adapt to the movement of equipment on the construction site by reconstructing the digital twin baseline; through theoretical simulation and feature coupling decision, non-human environmental interference, such as wind blowing grass and changes in light and shadow, is eliminated from the physical principle level, which significantly reduces the false alarm rate.

[0016] Example 2: Using physically based rasterization rendering technology, a baseline 3D projection image and baseline depth map with no human interference and consistent with the camera's perspective are generated. Based on the construction schedule, the coordinates of legal equipment locations and tower material stacking points at the current construction stage are analyzed, and the BIM model layout is updated. Combining GIS geographic information and the current time point, the solar azimuth and elevation angles are calculated to generate scene shadow masks. A virtual camera view frustum is constructed using the camera's intrinsic and extrinsic parameters. Combined with the scene shadow mask and current weather visibility parameters, the updated BIM model is rasterized and rendered, outputting a baseline 3D projection image containing RGB color channels and a baseline depth map containing distance information.

[0017] This embodiment further refines the generation process of the reference three-dimensional projection image and the reference depth map, especially how to handle the dynamically changing construction environment; Regarding dynamic BIM model updates, the legal equipment location coordinates refer to the spatial coordinates of construction machinery, such as cranes, excavators, materials, or tower materials, that are allowed to exist in the current time period according to the daily construction schedule. Implementation logic: The processor parses the construction schedule. If the schedule shows that tower erection work will be carried out today, the crane model in the BIM model is moved to the work point specified in the schedule. If the schedule shows that materials are arriving on site, the tower material model is generated at the stacking point on the GIS ground. This step ensures that the baseline is not static, but dynamically evolves with the working conditions. The implementation logic for ambient lighting and meteorological simulation combines GIS geographic information (latitude and longitude) with the current time to accurately calculate the solar azimuth and altitude angles, and uses shadow mapping technology to generate scene shadow masks. Meteorological parameter mapping: To address the black-box problem of adjusting rendering transmittance, this embodiment uses an exponential fog effect model to map meteorological visibility parameters to pixel-level operations; let the visibility be... meters, then the extinction coefficient ,in, The dimensionless atmospheric scattering constant is taken as approximately 3.912 according to Koschmieder's law, and its depth value in the Z-buffer is read. Calculate the final output color : ; Among them, the above formula is for The R, G, and B color channels are calculated independently. For rasterized colors, The current ambient light color is calculated using the sky region shown below; this color value... The average RGB pixel value of the sky region in the real-time monitoring video frame is obtained by calculating the sky region, which is the area above the image with a preset proportion or the area semantically segmented as the sky. This step ensures that the reference image has the same contrast attenuation characteristics as the real camera under hazy or dusty weather. View frustum construction and rasterization rendering, implementation logic: utilizing the camera's intrinsic parameter matrix. and extrinsic parameter matrix The system constructs the view frustum of the virtual camera; combines the updated BIM model and scene shadow mask for rasterization rendering; outputs an RGB image that reflects the ideal visual effect under the current lighting and layout; and records the distance of each pixel from the camera for subsequent geometric occlusion judgment. This embodiment introduces a construction schedule to drive BIM updates and uses a physics-based fog effect rendering system. The system can automatically adapt to legitimate scene changes and weather interference, avoiding background modeling failure caused by drastic environmental changes.

[0018] Example 3: The process involves calculating the pixel and depth differences between real-time monitoring video frames and a reference 3D projection image to generate a real-time monitoring differential feature map containing unexpected changes. This includes: performing semantic segmentation on the real-time monitoring video frames and the reference 3D projection image to obtain semantic label maps; performing consistency comparison on the semantic label maps to identify discrepancy regions with inconsistent semantic categories, and removing pixel regions belonging to legitimate construction machinery movement from the discrepancy regions based on the construction plan; extracting pixel-level and depth-level residual values ​​from the remaining discrepancy regions, and stitching them together to generate the real-time monitoring differential feature map.

[0019] This embodiment details how to generate a real-time monitoring differential feature map, especially how to use semantic segmentation to eliminate legitimate interference; Dual-path semantic segmentation: The semantic label map refers to the mapping map that assigns a specific category label to each pixel in the image, such as ground, machinery, vegetation, and people. Implementation logic: A lightweight semantic segmentation network, such as DeepLabV3+ or BiSeNet, is used to infer from real-time monitoring video frames and reference 3D projection images to generate two semantic label maps. Semantic comparison and legality removal: The logic is as follows: A consistency comparison is performed on two semantic label images to identify areas with inconsistent categories, such as ground in the baseline image and an unknown object in the real-time image. Key filtering: Based on the construction plan, if a discrepancy area is identified as construction machinery and the location falls within the planned permitted operating range, the pixel area is marked as legal and removed from the discrepancy area. However, priority verification is performed before removal: if the semantic label of the discrepancy area contains the category of "person," even if it is within the legal machinery operating range, the removal operation is not performed to prevent false alarms caused by personnel being obscured or overlapping by machinery. This effectively prevents false alarms for legally operating vehicles. Feature vector generation, implementation logic: For the remaining difference regions, i.e., the unexpected regions that cannot be explained, extract the L2 norm of the RGB difference of the pixel-level residual values ​​and the depth-level residual values, mainly for three-dimensional objects; perform Min-Max normalization on the pixel-level residual values ​​and the depth-level residual values ​​respectively, map them to a unified dimensionless interval of 0,1, and then concatenate these values ​​to form a real-time monitoring differential feature map; this vector essentially describes the semantic category, spatial location and geometric scale of the target; This embodiment introduces semantic segmentation and construction plan filtering, enabling the system to understand the content of the screen and accurately distinguish between legitimate moving targets and illegal intrusion targets.

[0020] Example 4: The process of generating a virtual 3D human body model in the 3D space corresponding to the reference 3D projection image includes: reading the safety distance values ​​in the power safety work regulations and delineating the 3D space area that is prohibited from entering in the BIM model; generating a set of virtual 3D human body models based on human skeletal kinematic constraints and placing the virtual 3D human body models on the ground grid in the 3D space area that is prohibited from entering; and simulating the occlusion relationship and motion posture of the model under the current camera viewpoint based on the perspective projection relationship of the camera.

[0021] This embodiment illustrates the specific strategy for generating a virtual three-dimensional human body model in virtual space, demonstrating the constraining effect of physical rules; Spatial constraints: Delineating the hazardous area, the three-dimensional spatial region where entry is prohibited refers to the three-dimensional envelope space formed by the minimum air insulation distance that must be maintained around live parts as defined in the electrical safety work regulations; Parameter source: Read from the regulations database, for example, for 110kV lines, setting safety distances. In the BIM model, the area is expanded from the energized part as the center. Distance creates a restricted area; Skeletal Constraints: Generating Virtual Models. Implementation Logic: The system does not simply paste images, but generates 3D mesh models. Based on human skeletal kinematic constraints, such as the knee joint only being able to bend backward and the stride not exceeding a certain proportion of height, a set of virtual 3D human models are generated. These models are placed on the ground mesh within the restricted area. Using the camera's perspective projection relationship, the size, perspective distortion, and occlusion relationship of the model under the current viewpoint are calculated. This embodiment generates a virtual model through physical constraints. The system can not only determine whether there are people present, but also whether a real security threat has been posed, such as a distance violation, thus realizing a compliant interpretation of the warning.

[0022] Example 5: Generating simulated invasion feature maps that characterize human invasion patterns through physical rendering includes: rendering a scene with virtual 3D human models as a simulated disturbed image and calculating the pixel difference between the simulated disturbed image and the reference 3D projected image; extracting texture and contour features of pixel change areas caused only by virtual 3D human models; generating corresponding simulated invasion feature maps for each model in a set of virtual 3D human models, thereby constructing a candidate set of simulated invasion feature maps.

[0023] This embodiment describes how to extract purely simulated intrusion feature maps from a simulated scenario; Simulated disturbed image rendering, implementation logic: After placing a virtual 3D human body model into the scene, perform local re-rendering to obtain a simulated disturbed image; this image represents what the surveillance screen should look like if someone were to intrude. Pure feature extraction, implementation logic: Calculate the pixel difference between the simulated disturbed image and the reference 3D projection image; since the two differ only at the human body model, the background completely cancels out, thus extremely clean texture features, simulated texture and contour features of the human body surface, and edge shape of the human body can be extracted; these features constitute the simulated intrusion feature map; it is a perfect reference answer; This step generates a high signal-to-noise ratio matching template, providing a precise standard answer for subsequent feature coupling and avoiding the scenario mismatch problem caused by directly using public dataset samples.

[0024] Example 6: The calculation of the spatial overlap between the real-time monitoring differential feature map and the simulated intrusion feature map in the same pixel coordinate system includes: using the Sobel operator to extract the edge gradient magnitude features of the region corresponding to the pixel-level residual numerical channel in the real-time monitoring differential feature map; extracting the geometric contour features of the virtual three-dimensional human model in the simulated intrusion feature map; calculating the intersection-union ratio (IoU) of the edge gradient magnitude features and the geometric contour features of each feature map in the candidate set of the simulated intrusion feature map in the pixel coordinate system; and selecting the largest IoU value as the spatial overlap.

[0025] This embodiment details the specific calculation method for spatial overlap, namely morphological fingerprint alignment. To address the data mismatch issue between gradient features and geometric contours, which cannot be directly used for IoU calculation, this embodiment employs the following processing logic: Edge gradient extraction and binarization, implementation logic: For the image region ROI corresponding to the real-time monitoring differential feature map, the Sobel operator is applied to calculate the gradient magnitude of each pixel: ; in, and Representing pixels The horizontal and vertical gradient values ​​at the given location; the key step uses Otsu's method to adaptively calculate the gradient threshold. This transforms a continuous gradient map into a discrete binary edge mask. Furthermore, a morphological closing operation is performed on the Mreal to fill the internal voids caused by clothing texture or uneven lighting, thus forming a solid connected region: ; This step transforms the blurry edges in reality into a binary shape with a computable area; Geometric contour extraction and projection, implementation logic: The virtual human 3D mesh model in the simulated intrusion feature map is projected using a camera projection matrix. Projecting onto a 2D pixel plane generates a binarized virtual silhouette mask. The internal filling state of this silhouette is directly preserved without edge extraction, resulting in a virtual solid mask. The area covered by the virtual human projection is 1, and the rest is 0; The Cross-Union Ratio (IoU) is calculated using the formula defined above, based on the two binary masks mentioned above. ; in, This indicates that within the pixel coordinate system of the real-time monitoring differential feature map, for all pixel coordinates... The binary mask values ​​of 0 or 1 are iterated and accumulated; Only when the edges of real objects are distributed With simulated human silhouette The IoU value will be high only when there is a high degree of overlap in the pixel coordinate system; This embodiment solves the mathematical logic defect that continuous gradient fields and discrete geometric contours cannot be directly subjected to set operations by introducing a binarized mask transformation, thus ensuring the feasibility of spatial overlap calculation.

[0026] Example 7: The processor is also configured to perform a temporal causality verification step: extract optical flow field data from multiple consecutive real-time monitoring video frames and calculate the displacement vector of the difference region; obtain the simulated motion trajectory vector of the virtual 3D human model within a preset time window and calculate the cosine similarity between the displacement vector and the simulated motion trajectory vector; if the cosine similarity is greater than the preset motion law threshold, the temporal causality verification is confirmed to have passed; if the cosine similarity is less than or equal to the motion law threshold, the temporal causality verification is confirmed to have failed.

[0027] This embodiment introduces a temporal causality verification step; to address the dimensionality mismatch issue between directly comparing the optical flow field data vector field and the single vector of the simulated motion trajectory, this embodiment employs the following processing logic: Real-world optical flow field extraction and aggregation: Implementation logic: Extract multiple consecutive frames, such as 5-10 frames, from real-time monitored video frames, and use the Lucas-Kanade optical flow method to calculate each feature point within the difference region. instantaneous velocity vector Key step: Vector aggregation. To reduce the dimensionality of the field data into feature vectors comparable to the trajectory, a weighted average of all effective optical flow vectors within the difference region is calculated to obtain the actual displacement vector. : ; in, This represents the total number of pixels within the difference region. For the first The instantaneous velocity vector of each feature point; denominator This represents the sum of weights of all feature points within the difference region, used for normalizing the velocity vector; The confidence weights based on gradient magnitude are calculated using the following formula: ; in, For the first Gradient magnitude of each pixel For example, the preset minimum value. This is used to prevent invalid denominator calculations due to the sum of gradients being zero; Theoretical trajectory similarity calculation, parameter definition: simulated motion trajectory vector This refers to calling the human kinematics model generated in Example 4, within the same time window. Within the virtual human body, the displacement vector of the center of mass in the two-dimensional image plane is defined by the formula: Calculate the cosine similarity between the aggregated real vector and the simulated vector: ; in, For the actual displacement vector With simulated motion trajectory vector The angle between them The L2 norm of a vector; Logical criterion: If the cosine similarity is greater than the threshold of the motion law. For example, 0.8 indicates that the overall movement trend of real objects conforms to the physical directionality of human walking, confirming that the temporal causality check has passed; This embodiment solves the logical loophole that point-to-point optical flow fields cannot be directly compared with macroscopic trajectories through a vector aggregation step, effectively distinguishing between the overall moving human body and the optical flow vector and the stationary disorderly oscillating vegetation that approaches 0.

[0028] Example 8: The determination of whether a genuine person has entered the premises and the triggering of an alarm signal include: generating an alarm signal containing the coordinates and time of the intrusion location only when the spatial overlap is greater than the spatial consistency threshold and the temporal causality check passes, and sending it to the on-site audible and visual alarm; if the spatial overlap is less than or equal to the spatial consistency threshold, or the temporal causality check fails, then the real-time monitoring differential feature map is determined to be environmental background noise, no alarm signal is generated, and the current log is recorded as a false alarm event.

[0029] This embodiment defines the final comprehensive decision logic, reflecting the high-precision principle of "better to have none than to have something of poor quality"; Alarm triggering logic, AND gate logic, implementation logic, the system executes a strict logical AND operation: ; That is, only when the spatial overlap is greater than a preset threshold, conforms to the physical morphological characteristics of the human body, and passes the temporal causality check, conforms to the laws of human movement, will the processor generate an alarm signal containing the intrusion location coordinates and time calculated by BIM coordinate back projection, and send it to the on-site audible and visual alarm through the communication interface. False alarm suppression and logging implementation logic: If any condition is not met, for example, a static human-like false alarm that fails the temporal verification even though the spatial overlap meets the standard, or a non-human interference that does not meet the spatial overlap standard even though there is displacement, the system will determine it as irrelevant background noise. At this time, the system will not trigger an alarm, but will record the feature vector of the event in the log and mark it as a false alarm event for subsequent adaptive threshold adjustment or model iteration of the system. This embodiment ensures extremely high system reliability through a dual-threshold logic design, minimizing the interference of false alarms on construction progress, while the logging function provides data support for continuous system optimization.

[0030] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A visual early warning system for preventing unauthorized personnel from entering power transmission line construction sites, characterized in that, include: The processor, memory, and communication interface; the memory stores a computer program, and the processor executes the computer program to perform the following steps: Data acquisition and benchmark reconstruction steps: acquire BIM model data of transmission line construction, GIS geographic information, construction schedule and camera internal and external parameters, calculate solar altitude angle based on current time, and use rasterization rendering technology to render and generate a benchmark 3D projection image and benchmark depth map that are consistent with the camera's viewpoint and free from human interference. Reality difference extraction steps: Real-time monitoring video frames of the construction site are acquired by the camera, the pixel differences and depth differences between the real-time monitoring video frames and the reference three-dimensional projection image are calculated, and a real-time monitoring differential feature map containing unexpected changes is generated. Theoretical simulation deduction steps: Based on the human skeletal kinematic parameters in the preset risk parameter library, a virtual three-dimensional human model is generated in the three-dimensional space corresponding to the benchmark three-dimensional projection image, and a simulated invasion feature map representing the human invasion pattern is generated through physical rendering. Feature coupling decision step: Calculate the spatial overlap between the real-time monitoring differential feature map and the simulated intrusion feature map in the same pixel coordinate system; If the spatial overlap is greater than a preset spatial consistency threshold, it is determined that there is a real person mistakenly entering the space and an alarm signal is triggered; if the spatial overlap is less than or equal to the spatial consistency threshold, it is determined to be environmental noise and the current monitoring status is maintained.

2. The visual early warning system for preventing unauthorized personnel from entering during power transmission line construction, as described in claim 1, is characterized in that... The process of using rasterization rendering technology to generate a reference 3D projection image and reference depth map that are consistent with the camera's perspective and free from human interference specifically includes: parsing the legal equipment location coordinates and tower material stacking point coordinates of the current construction stage according to the construction schedule, and updating the BIM model layout; calculating the solar azimuth and elevation angles by combining the GIS geographic information and the current time point, and generating a scene shadow mask; constructing a virtual camera view frustum using the camera's intrinsic and extrinsic parameters, and performing rasterization rendering on the updated BIM model by combining the scene shadow mask, ambient light color extracted from the sky area of ​​the real-time monitoring video frame, and current meteorological visibility parameters, and outputting the reference 3D projection image containing RGB color channels, and the reference depth map containing distance information.

3. The visual early warning system for preventing unauthorized personnel from entering power transmission line construction sites according to claim 1, characterized in that, The step of calculating the pixel and depth differences between the real-time monitoring video frame and the reference 3D projection image to generate a real-time monitoring differential feature map containing unexpected changes specifically includes: performing semantic segmentation on the real-time monitoring video frame and the reference 3D projection image respectively to obtain semantic label maps; performing consistency comparison operations on the semantic label maps to identify difference regions with inconsistent semantic categories, and based on the construction plan table, removing pixel regions belonging to legitimate construction machinery movement from the difference regions; extracting pixel-level residual values ​​and depth-level residual values ​​from the remaining difference regions, and generating the real-time monitoring differential feature map by stitching them together.

4. The visual early warning system for preventing unauthorized personnel from entering during power transmission line construction, as described in claim 1, is characterized in that... The step of generating a virtual three-dimensional human body model in the three-dimensional space corresponding to the reference three-dimensional projection image specifically includes: reading the safety distance value in the power safety work regulations, delineating the prohibited three-dimensional space area in the BIM model; generating a set of virtual three-dimensional human body models based on human skeletal kinematic constraints, and placing the virtual three-dimensional human body models on the ground grid in the prohibited three-dimensional space area, and simulating their occlusion relationship and movement posture under the current camera view according to the camera perspective projection relationship.

5. A visual early warning system for preventing unauthorized personnel from entering power transmission line construction sites, as described in claim 4, is characterized in that... The process of generating a simulated invasion feature map representing the morphology of human invasion through physical rendering specifically includes: rendering a scene with a virtual three-dimensional human model as a simulated disturbed image, and calculating the pixel difference between the simulated disturbed image and the reference three-dimensional projection image; extracting the texture and contour features of the pixel change area caused only by the virtual three-dimensional human model, and generating the simulated invasion feature map.

6. A visual early warning system for preventing unauthorized personnel from entering power transmission line construction sites according to claim 1, characterized in that, The calculation of the spatial overlap between the real-time monitoring differential feature map and the simulated intrusion feature map in the same pixel coordinate system specifically includes: extracting the edge gradient magnitude features of the corresponding region of the real-time monitoring differential feature map using the Sobel operator; extracting the geometric contour features of the virtual three-dimensional human model in the simulated intrusion feature map; calculating the intersection-over-union (IoU) value of the edge gradient magnitude features and the geometric contour features in the pixel coordinate system, and using the IoU value as the spatial overlap.

7. A visual early warning system for preventing unauthorized personnel from entering power transmission line construction sites, as described in claim 6, is characterized in that... The processor is further configured to perform a temporal causality verification step: extracting optical flow field data from multiple consecutive frames of the real-time monitoring video frames, calculating displacement vectors in the difference regions; obtaining the simulated motion trajectory vector of the virtual three-dimensional human model within a preset time window, and calculating the cosine similarity between the displacement vector and the simulated motion trajectory vector; if the cosine similarity is greater than a preset motion law threshold, then the temporal causality verification is confirmed to have passed; if the cosine similarity is less than or equal to the motion law threshold, then the temporal causality verification is confirmed to have failed.

8. A visual early warning system for preventing unauthorized personnel from entering power transmission line construction sites, as described in claim 7, is characterized in that... The determination of whether a real person has entered the premises and triggering an alarm signal specifically includes: generating an alarm signal containing the coordinates and time of the intrusion location only when the spatial overlap is greater than the spatial consistency threshold and the temporal causality check passes, and sending it to the on-site audible and visual alarm; if the spatial overlap is less than or equal to the spatial consistency threshold, or the temporal causality check fails, then the real-time monitoring differential feature map is determined to be environmental background noise, no alarm signal is generated, and the current log is recorded as a false alarm event.

Images