Substation visual ranging method, system, device, and medium

Abstract

The application relates to the technical field of substation operation, and discloses a substation visual distance measurement method, system, device and medium, the method constructs a real-time dense point cloud model of each sampling point through a multi-view image sequence of a target monitoring area at continuous sampling points, so that the real-time and authenticity of the dense point cloud model are ensured, dynamic target segmentation is carried out on scene point clouds of the real-time dense point cloud model, point cloud clusters of dynamic targets can be accurately extracted from a complex substation background, the continuous three-dimensional motion trajectory of the dynamic target is obtained through cross-frame association and state estimation of the dynamic target, the robustness of target identification and tracking is significantly improved, distance measurement target tracking loss or drift is avoided, and the spatial distance between the dynamic target and a preset distance measurement observation point is solved in real time by using the three-dimensional motion trajectory of the dynamic target, so that the distance measurement precision is improved, and the false detection rate and the safety distance misjudgment rate are reduced.

Description

Technical Field

[0001] This invention relates to the field of substation operation technology, and in particular to a visual ranging method, system, equipment and medium for substations. Background Technology

[0002] With the rapid development of intelligent and unmanned operation and maintenance in power systems, substations, as key nodes in the power grid, are increasingly important for the safety management and operational standardization monitoring of their equipment and personnel. Traditional substation safety monitoring mainly relies on video surveillance and manual inspections, which makes it difficult to accurately locate, measure distances in real time, and analyze the behavior of dynamic targets such as personnel and construction machinery. This results in limitations such as large monitoring blind spots, delayed early warnings, and the inability to quantify safe distances.

[0003] In recent years, computer vision-based 3D perception technology, especially dense point cloud technology generated through visual reconstruction, has provided new solutions for the three-dimensional and high-precision safety monitoring of substations.

[0004] However, due to the complex environment of substations, there are a large number of structured power equipment (such as transformers, frames, and insulator strings) in the background. Their geometry is easily confused with moving targets (such as personnel and machinery) in the point cloud, resulting in poor accuracy of visual ranging and a high false detection rate. In addition, there are situations where multiple targets occlude each other and their movement trajectories intersect, which also makes it very difficult to track targets continuously and stably. This can easily lead to the loss or drift of the ranging target, and thus cause misjudgment of the safe distance. Summary of the Invention

[0005] In view of this, the present invention provides a visual ranging method, system, equipment and medium for substations, which solves the technical problems of poor accuracy, high false detection rate and loss or drift of the ranging target, which leads to misjudgment of safe distance.

[0006] The first aspect of this invention provides a visual ranging method for substations, comprising:

[0007] A multi-view image sequence of the target monitoring area in the substation at continuous sampling points is obtained, and real-time three-dimensional reconstruction is performed based on the multi-view image sequence to obtain a real-time dense point cloud model of each sampling point.

[0008] For each sampling point, dynamic target segmentation is performed on the scene point cloud of the real-time dense point cloud model to obtain the point cloud cluster of dynamic targets for each sampling point;

[0009] Based on the point cloud clusters of the dynamic targets at each sampling point, cross-frame association and state estimation are performed on the dynamic targets to obtain the three-dimensional motion trajectory of the dynamic targets.

[0010] Based on the three-dimensional motion trajectory of the dynamic target, the spatial distance between the dynamic target and the preset ranging observation point is calculated in real time.

[0011] In one embodiment, the step of acquiring a multi-view image sequence of the target monitoring area within the substation at continuous sampling points, and performing real-time 3D reconstruction based on the multi-view image sequence to obtain a real-time dense point cloud model for each sampling point includes:

[0012] The multi-view image sequence is obtained by simultaneously acquiring image sequences of the target monitoring area from multiple perspectives using a multi-view camera.

[0013] The multi-view image sequence is time-synchronized and spatially registered to obtain a registered multi-view image sequence.

[0014] Feature extraction is performed on the registered multi-view image sequence, and cross-view matching relationships are established based on the extracted features to obtain feature point matching pairs between each view.

[0015] By utilizing multi-view geometric rules and combining feature point matching pairs between the various viewpoints, the camera spatial pose under each viewpoint is determined.

[0016] Based on the feature point matching pairs between the various viewpoints and the camera spatial pose, the three-dimensional coordinates of the scene feature points are recovered by triangulation, and the depth information of the multi-view image sequence is obtained. The depth information of each viewpoint is fused with the three-dimensional coordinates to obtain a dense depth map.

[0017] The dense depth maps of each sampling point are incrementally registered and fused in a unified coordinate system to generate an initial real-time dense point cloud model for each sampling point.

[0018] Occlusion compensation and surface interpolation optimization are performed on the initial real-time dense point cloud model to obtain the real-time dense point cloud model for each sampling point.

[0019] In one embodiment, the step of performing dynamic target segmentation on the scene point cloud of the real-time dense point cloud model for each sampling point to obtain a point cloud cluster of dynamic targets for each sampling point includes:

[0020] Temporal registration and background motion compensation are performed on the scene point cloud of the real-time dense point cloud model of each pair of adjacent sampling points to obtain the compensated scene point cloud.

[0021] Based on motion consistency constraints, all motion-salient point cloud regions are extracted from the compensated scene point cloud to construct an initial dynamic candidate point cloud region set.

[0022] For each candidate point cloud region in the initial dynamic candidate point cloud region set, edge refinement and connectivity constraints are applied to the candidate point cloud regions to obtain refined candidate point cloud regions.

[0023] Deep mutation detection is performed on the point cloud within the refined candidate point cloud region to obtain the boundary point set of the refined candidate point cloud region;

[0024] Based on the dense point cloud corresponding to the refined candidate point cloud region, and combined with a preset semantic feature extraction network, the device semantic category of the refined candidate point cloud region is determined.

[0025] The motion features of the refined candidate point cloud regions are obtained. Based on the motion features, the boundary point set, and the device semantic category, candidate point cloud regions of dynamic targets are selected from all the refined candidate point cloud regions and used as masks for the dynamic targets.

[0026] The point cloud clusters of the dynamic target are extracted from the real-time dense point cloud model using the mask of the dynamic target.

[0027] In one embodiment, the step of performing cross-frame association and state estimation on the dynamic target based on the point cloud clusters of the dynamic target at each of the sampling points to obtain the three-dimensional motion trajectory of the dynamic target includes:

[0028] The three-dimensional geometric features of the point cloud clusters of dynamic targets at each sampling point are extracted, and the motion direction and displacement changes of dynamic targets at adjacent sampling points are estimated using the optical flow method.

[0029] Based on the motion direction and displacement changes of the dynamic target, and combined with the three-dimensional geometric features, the cross-frame target association relationship of the dynamic target at each sampling point is determined;

[0030] Based on the cross-frame target association, the state of the dynamic target is recursively estimated using Kalman filtering, and the observation results of the current sampling point are fused to determine the optimal estimated position of the dynamic target at the current sampling point.

[0031] Based on the optimal estimated position of the dynamic target, a continuous three-dimensional trajectory between each sampling point is constructed to obtain the three-dimensional motion trajectory of the dynamic target.

[0032] In one embodiment, the step of calculating the spatial distance between the dynamic target and a preset ranging observation point in real time based on the three-dimensional motion trajectory of the dynamic target includes:

[0033] The three-dimensional motion trajectory of the dynamic target is mapped to the preset ranging observation point in a unified coordinate system, and the three-dimensional motion trajectory of the dynamic target is discretized and sampled in the unified coordinate system to obtain multiple discrete motion points of the dynamic target.

[0034] For each of the moving discrete points, the Euclidean distance between the moving discrete point and the preset ranging observation point is calculated to determine the spatial distance between the dynamic target and the preset ranging observation point.

[0035] In one embodiment, the step of calculating the spatial distance between the dynamic target and a preset ranging observation point in real time based on the three-dimensional motion trajectory of the dynamic target further includes:

[0036] Obtain the distance measurement error impact data of the current sampling point within the substation;

[0037] Based on the pre-built error impact model and combined with the ranging error impact data, the ranging error compensation amount for the current sampling point is determined; wherein, the error impact model is a physical driving model built based on the mapping relationship between the ranging error impact data and its corresponding ranging error compensation amount;

[0038] The spatial distance is corrected based on the distance measurement error compensation amount to obtain the corrected spatial distance.

[0039] In one embodiment, the step of correcting the spatial distance based on the ranging error compensation amount to obtain the corrected spatial distance further includes:

[0040] Obtain the reference spatial distance between the dynamic target and the preset ranging observation point, and calculate the compensation deviation value based on the corrected spatial distance and the reference spatial distance;

[0041] When the compensation deviation value exceeds the preset safety threshold, the parameters of the error influence model are adjusted with the goal of minimizing the compensation deviation, the error influence model is updated, and the pre-built error influence model is re-executed based on the updated error influence model. Combined with the ranging error influence data, the ranging error compensation amount of the current sampling point is determined until the compensation deviation value does not exceed the preset safety threshold, and the final corrected spatial distance is output.

[0042] Secondly, the present invention also provides a substation visual ranging system, comprising:

[0043] The point cloud reconstruction module is used to acquire multi-view image sequences of the target monitoring area in the substation at continuous sampling points, and to perform real-time three-dimensional reconstruction based on the multi-view image sequences to obtain a real-time dense point cloud model for each sampling point.

[0044] The target point cloud segmentation module is used to perform dynamic target segmentation on the scene point cloud of the real-time dense point cloud model for each sampling point, so as to obtain the point cloud cluster of the dynamic target for each sampling point;

[0045] The target tracking module is used to perform cross-frame association and state estimation of the dynamic target based on the point cloud clusters of the dynamic target at each sampling point, so as to obtain the three-dimensional motion trajectory of the dynamic target.

[0046] The target ranging module is used to calculate the spatial distance between the dynamic target and the preset ranging observation point in real time based on the three-dimensional motion trajectory of the dynamic target.

[0047] Thirdly, the present invention also provides an electronic device, the electronic device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the substation visual ranging method as described in the first aspect.

[0048] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the steps of the substation visual ranging method as described in the first aspect.

[0049] As can be seen from the above technical solutions, this invention constructs a real-time dense point cloud model for each sampling point by using multi-view image sequences of the target monitoring area at continuous sampling points, thereby ensuring the real-time performance and realism of the dense point cloud model. Furthermore, it performs dynamic target segmentation on the scene point cloud of the real-time dense point cloud model, enabling accurate extraction of point cloud clusters of dynamic targets from complex substation backgrounds. By performing cross-frame association and state estimation on dynamic targets, it obtains the continuous three-dimensional motion trajectory of the dynamic targets, significantly improving the robustness of target recognition and tracking, avoiding target tracking loss or drift, and utilizing the three-dimensional motion trajectory of the dynamic targets to calculate the spatial distance between the dynamic targets and preset ranging observation points in real time, thereby improving ranging accuracy and reducing false detection rate and safe distance misjudgment rate. Attached Figure Description

[0050] 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is an application environment diagram of a visual ranging method for substations provided in an embodiment of the present invention;

[0052] Figure 2 A flowchart of a visual ranging method for substations provided in an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of the hardware connection structure provided in an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of the structure of a substation visual ranging system provided in an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0056] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0057] With the development of intelligent operation and maintenance and unmanned management of substations, the monitoring of safe distances between operators and live equipment has placed higher demands on real-time performance, accuracy, and reliability. Visual 3D perception, especially dense point cloud technology, is considered a potential direction for achieving high-precision safety monitoring. However, existing research still has systemic shortcomings when dealing with the complexities, strong interference, and high safety standards of special industrial scenarios in substations, making it difficult to engineering-scale laboratory algorithms. These problems stem from multi-level deficiencies in perception mechanisms and system architecture, becoming key bottlenecks restricting the practical application of high-precision visual ranging technology.

[0058] Specifically, existing technical solutions have the following technical problems when facing real-world substation scenarios:

[0059] First, in the extremely complex and dynamic substation environment, existing dynamic target segmentation and tracking methods suffer from inherent fragility in perception capabilities, failing to provide stable, continuous, and accurate target location awareness. Substations are not regular, open laboratories or structured roads; their environment is highly coupled with dense equipment, structural complexity, random occlusion, and electromagnetic interference. Current mainstream solutions, whether relying on fusion methods triggered by offline static maps and 2D images, or traditional point cloud segmentation algorithms based on simple geometric features, all rely on overly idealized assumptions in their perception models. They struggle to address:

[0060] Problems with continuous target perception under dense occlusion: When people move through transformers, circuit breakers, and structures, the target is frequently and extensively occluded, making the tracking chain based on continuous appearance or motion extremely prone to breakage.

[0061] Accurate separation in complex backgrounds: The metal boom of construction machinery and the background steel structure have similar features in point clouds. Traditional clustering algorithms are prone to confusing the two, resulting in a large number of missegments.

[0062] Adaptation problems in dynamic environments: Sudden changes in illumination, alternation of day and night, and rainy / foggy weather cause drastic fluctuations in point cloud quality, severely degrading the performance of algorithm models with fixed parameters. The direct consequence is that the target position sequence output by the perception system contains a large number of "jumps," "losses," and "identity swaps," causing any subsequent ranging calculations to lose an accurate and reliable data source, and the foundation of the entire system's reliability collapses.

[0063] Second, in the uniquely strong interference environment of substations, the lack of a systematic and modelable ranging error compensation system prevents the leap from "relative measurement" to "absolute accuracy." While visual ranging may achieve millimeter-level relative accuracy after laboratory calibration, the errors introduced in substation environments are multidimensional, time-varying, and non-Gaussian. Existing technologies either completely ignore these errors or can only compensate for static system errors introduced by sensor factory calibration, remaining ineffective against environmentally coupled dynamic errors. This includes:

[0064] (1) Physical environment interference error: strong alternating electromagnetic field causes reading drift of camera CMOS sensor or lidar photoelectric element; optical "multipath effect" caused by mirror reflection of large metal surface, producing non-real false points;

[0065] (2) Errors due to changes in the propagation medium: Local temperature gradients and oil-gas mixtures within the substation cause uneven air refractive index, resulting in unexpected deflection of light propagation;

[0066] (3) Errors introduced by target characteristics: Differences in the reflectivity of personnel's clothing and safety helmets to lasers, and motion blurring of point clouds caused by the rapid movement of mechanical moving parts. These errors are not random noise, but systematic deviations with physical causes. If they are not identified and compensated, and the original point cloud coordinates are used directly for geometric ranging, the results will have unpredictable and potentially excessive inherent deviations, rendering high-precision measurements meaningless and unable to provide a reliable basis for safety early warning decisions.

[0067] Third, existing technical solutions lack integrated system design oriented towards practical engineering applications, exhibiting inherent deficiencies in deployability, adaptability, and long-term stability. A monitoring system capable of long-term service in a substation must be a technically feasible, engineeringally usable, and economically maintainable organic whole. Existing research largely focuses on improving the performance of individual algorithm modules, lacking system-level considerations.

[0068] (1) The contradiction between computational complexity and real-time performance: Generating dense point clouds, running complex segmentation networks, and performing high-dimensional error compensation require extremely high computing power, making it difficult to achieve real-time frame rate processing on edge computing devices deployed on-site.

[0069] (2) The system is rigid and the scenario changes are contradictory: the fixed parameters and models cannot adapt to the long-term environmental changes brought about by the adjustment of substation equipment layout, the commissioning of new equipment, and the changing seasons. The system performance degrades with the running time.

[0070] (3) Conflict between deployment and maintenance costs: Schemes that rely on high-precision offline calibration (such as setting up multiple ball targets) are almost impossible to implement and reproduce on site, resulting in difficult deployment, long calibration cycles, and high maintenance costs. These problems have kept many advanced algorithms at the paper and demonstration stage, and they cannot form a standardized industrial product that can run stably 24 / 7 without interruption and requires no or low maintenance.

[0071] Therefore, in order to solve the above-mentioned technical problems, embodiments of this application provide a visual ranging method for substations, which can be applied to, for example... Figure 1 In the application environment shown, terminal 101 communicates with server 102 via a network. A data storage system can store the data that server 102 needs to process. The data storage system can be integrated onto server 102 or placed on a cloud or other network server. Terminal 101 or server 102 executes a substation visual ranging method, which includes: acquiring a multi-view image sequence of a target monitoring area within the substation at continuous sampling points; performing real-time 3D reconstruction based on the multi-view image sequence to obtain a real-time dense point cloud model for each sampling point; for each sampling point, performing dynamic target segmentation on the scene point cloud of the real-time dense point cloud model to obtain a point cloud cluster of dynamic targets at each sampling point; performing cross-frame association and state estimation on the dynamic targets based on the point cloud clusters of dynamic targets at each sampling point to obtain the 3D motion trajectory of the dynamic targets; and calculating the spatial distance between the dynamic targets and preset ranging observation points in real time based on the 3D motion trajectory of the dynamic targets.

[0072] Terminal 101 can be, but is not limited to, various personal computers, laptops, smartphones, and tablets.

[0073] Server 102 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services.

[0074] like Figure 2 As shown, this application provides a visual ranging method for substations, which is applied to... Figure 1 Taking terminal 101 or server 102 as an example, the explanation includes the following steps S1 to S4. Wherein:

[0075] Step S1: Obtain multi-view image sequences of the target monitoring area within the substation at continuous sampling points, and perform real-time three-dimensional reconstruction based on the multi-view image sequences to obtain a real-time dense point cloud model for each sampling point.

[0076] Among them, the target monitoring area is the area of ​​live equipment in the substation that needs to be monitored. By deploying multi-view sensors (RGB cameras, depth cameras or laser modules) or drones carrying multi-view cameras to synchronously sample the target monitoring area, multiple original images or depth data from multiple perspectives can be obtained. After sampling at continuous sampling points, the multi-view images of each sampling point are combined into a multi-view image sequence.

[0077] Meanwhile, real-time 3D reconstruction of multi-view image sequences can obtain a real-time dense point cloud model for each sampling point, enabling millimeter-level 3D stereoscopic scanning of the substation site. This transforms the planar image into measurable 3D data, with each point carrying precise 3D coordinates (X / Y / Z), thus clearly revealing the 3D shape and position of equipment, personnel, and machinery.

[0078] Step S2: For each sampling point, perform dynamic target segmentation on the scene point cloud of the real-time dense point cloud model to obtain the point cloud cluster of the dynamic target at each sampling point.

[0079] Scene point cloud is a collection of point clouds in a real-time dense point cloud model. In complex industrial environments, background structures and target objects are often mixed together. In order to achieve accurate extraction of dynamic targets, dynamic target segmentation can be performed on scene point cloud to obtain the motion region that deviates significantly from the background motion pattern in each sampling point, thereby separating the real dynamic target for subsequent target localization, state analysis and autonomous shooting control.

[0080] Step S3: Based on the point cloud clusters of the dynamic target at each sampling point, perform cross-frame association and state estimation on the dynamic target to obtain the three-dimensional motion trajectory of the dynamic target.

[0081] In this process, after obtaining the point cloud clusters of the dynamic target at each sampling point, it is necessary to obtain the positional changes of the dynamic target in time. Therefore, this application performs cross-frame association and state estimation on the point cloud clusters of the dynamic target extracted at continuous time points, thereby forming a stable and continuous three-dimensional motion trajectory of the dynamic target.

[0082] Among them, cross-frame association establishes the consistency of target identity in the time dimension by matching the geometric center, bounding box overlap and local shape features of target point cloud clusters in adjacent frames; state estimation uses Kalman filtering to update motion state parameters such as target position, velocity and acceleration in real time.

[0083] Step S4: Based on the three-dimensional motion trajectory of the dynamic target, calculate the spatial distance between the dynamic target and the preset ranging observation point in real time.

[0084] The dynamic target can be personnel, equipment components, foreign objects, or other dynamically relevant objects to be observed. The preset ranging observation point is a visible light camera, depth camera, lidar, or a combination of these sensing units mounted on the UAV platform. By calculating the spatial distance between the dynamic target and the preset ranging observation point, the relative position of the UAV and the target can be precisely controlled, achieving centimeter-level safe distance warnings. After ranging is completed, the results are uploaded to the monitoring center, providing a basis for remote diagnosis and status monitoring. The monitoring center can then issue update commands to achieve system-level remote maintenance and strategy optimization.

[0085] It should be noted that the embodiments of this application construct a real-time dense point cloud model for each sampling point by using multi-view image sequences of the target monitoring area at continuous sampling points, thereby ensuring the real-time performance and realism of the dense point cloud model. Furthermore, dynamic target segmentation is performed on the scene point cloud of the real-time dense point cloud model, enabling accurate extraction of point cloud clusters of dynamic targets from complex substation backgrounds. By performing cross-frame association and state estimation on the dynamic targets, the continuous three-dimensional motion trajectory of the dynamic targets is obtained, significantly improving the robustness of target recognition and tracking, avoiding target tracking loss or drift, and utilizing the three-dimensional motion trajectory of the dynamic targets to calculate the spatial distance between the dynamic targets and preset ranging observation points in real time, thereby improving ranging accuracy and reducing false detection rate and safe distance misjudgment rate.

[0086] In some embodiments, acquiring a multi-view image sequence of a target monitoring area within a substation at continuous sampling points, and performing real-time 3D reconstruction based on the multi-view image sequence to obtain a real-time dense point cloud model for each sampling point includes: synchronously acquiring image sequences of the target monitoring area at multiple perspectives using a multi-view camera to obtain a multi-view image sequence; performing time synchronization and spatial registration on the multi-view image sequence to obtain a registered multi-view image sequence; extracting features from the registered multi-view image sequence, establishing cross-view matching relationships based on the extracted features to obtain feature point matching pairs between different perspectives; and utilizing multi-view... Geometric rules are used to determine the camera spatial pose of each viewpoint by combining feature point matching pairs. Based on the feature point matching pairs and camera spatial pose, the 3D coordinates of scene feature points are recovered through triangulation, and depth information of multi-view image sequences is obtained. The depth information and 3D coordinates of each viewpoint are fused to obtain a dense depth map. The dense depth maps of each sampling point are incrementally registered and fused in a unified coordinate system to generate an initial real-time dense point cloud model for each sampling point. Occlusion compensation and surface interpolation optimization are performed on the initial real-time dense point cloud model to obtain a real-time dense point cloud model for each sampling point.

[0087] This involves using a drone equipped with a multi-view camera to simultaneously sample the target monitoring area, capturing image sequences of the target monitoring area from multiple perspectives. Simultaneously, the onboard inertial measurement unit rapidly acquires the drone's position, angle, acceleration, and angular velocity data. The raw images or depth data acquired by each channel are denoted as follows:

[0088]

[0089] In the formula, M represents the raw image or depth data of the multi-view sensor channel m at sampling point t, where M is the number of multi-view sensor channels.

[0090] To ensure a consistent description of the spatial geometry, each channel is time-stamp aligned using a common clock or synchronization pulse, and basic preprocessing (brightness normalization, noise suppression, distortion correction) is performed on the original image, which can be formalized as follows:

[0091]

[0092] In the formula, This is the preprocessed image data, used as input for subsequent 3D reconstruction.

[0093] To construct the spatial geometry of the scene, real-time 3D reconstruction is performed based on a multi-view image sequence. Specifically, the pose change information output by the inertial measurement unit is combined to perform time synchronization and spatial registration on the multi-view image sequence; subsequently, camera calibration and distortion correction are performed on the multiple images to obtain distortion-free, standardized, and clear multiple images, eliminating computational errors caused by lens physical defects.

[0094] Based on this, stable features that do not change with lighting / angle are captured from the corrected multi-channel images, such as equipment corners, screws, insulator edges, and structural intersections. The same features captured by different cameras at different times are matched. For example, the transformer corner captured by the left camera and the same corner captured by the right camera in the previous frame are matched to establish a precise cross-view matching relationship, thereby determining the corresponding feature point matching pairs between multiple views.

[0095] Furthermore, by utilizing multi-view geometric rules (such as epipolar constraints and coplanar constraints) combined with feature point matching pairs between different viewpoints, the camera's spatial pose in 3D space is calculated, including six degrees of freedom and three angles: pitch, yaw, and roll. These pose parameters, together with the feature point matching pairs, constitute the input for triangulation. By minimizing the reprojection error and optimizing the 3D point coordinates, the 3D coordinates of these feature points in the real world are recovered, thus completing the initial reconstruction of the sparse point cloud.

[0096] Then, based on the disparity calculation and depth estimation model, the depth information of the multi-view image sequence is obtained. The depth information and the reconstructed sparse point cloud are then rotated to the same coordinate system, and then densification interpolation is performed to generate a high-density, high-precision dense depth map that covers the current shooting area.

[0097] Then, the dense depth maps of each sampling point are incrementally registered in a unified coordinate system.

[0098] The multi-frame local point cloud generated by the continuous flight and shooting of the drone is stitched frame by frame into a unified global coordinate system, and continuously superimposed and expanded to form a large scene point cloud covering the entire inspection area, thus obtaining the initial real-time dense point cloud model of each sampling point.

[0099] For partially occluded or less textured areas, a combination of sparse point cloud expansion and surface interpolation methods can be used to improve the continuity and integrity of the point cloud. Specifically, for occluded areas (where the equipment blocks the view), a sparse point cloud expansion algorithm is used to fill in the missing 3D structure based on the surrounding point cloud trend; for less textured areas (such as smooth insulator surfaces), a surface interpolation algorithm is used to smoothly fill the point cloud holes, making the surface more continuous and complete. The final output is a real-time dense 3D point cloud model that covers the entire inspection area, is hole-free, highly accurate, and complete. Each point has precise 3D coordinates and can be directly used for subsequent target segmentation and ranging. The dense point cloud can be represented as:

[0100]

[0101] In the formula, This represents the dense 3D point cloud of the scene reconstructed at that moment. The point cloud not only contains the location and spatial structure information of the equipment targets within the monitoring area, but also preserves the surface texture and local geometric details of the scene, providing a 3D environmental foundation for subsequent dynamic target segmentation, target localization, and autonomous image capture control.

[0102] This application employs 3D field-of-view analysis and synchronous acquisition design for the substation monitoring area, ensuring comprehensive coverage of critical operating areas and avoiding blind spots inherent in traditional single-view monitoring. It also guarantees the temporal and spatial alignment of multi-view images from the outset. With a rationally designed acquisition system, the optical information of dynamic targets (personnel, vehicles, equipment) within the monitoring area is fully and synchronously captured, providing a high-quality data foundation for subsequent high-precision 3D reconstruction.

[0103] In some embodiments, for each sampling point, dynamic target segmentation is performed on the scene point cloud of the real-time dense point cloud model to obtain a point cloud cluster of dynamic targets for each sampling point. This includes: performing temporal registration and background motion compensation on the scene point clouds of the real-time dense point cloud model for every pair of adjacent sampling points to obtain compensated scene point clouds; based on motion consistency constraints, extracting all motion-salient point cloud regions from the compensated scene point clouds to construct an initial dynamic candidate point cloud region set; and for each candidate point cloud region in the initial dynamic candidate point cloud region set, performing edge refinement and connectivity constraints on the candidate point cloud regions to obtain refined candidate points. The process involves: 1) Performing deep mutation detection on the point clouds within the refined candidate point cloud regions to obtain the boundary point set of the refined candidate point cloud regions; 2) Determining the device semantic category of the refined candidate point cloud regions based on the dense point clouds corresponding to the refined candidate point cloud regions, combined with a pre-defined semantic feature extraction network; 3) Obtaining the motion features of the refined candidate point cloud regions; 4) Selecting candidate point cloud regions for dynamic targets from all refined candidate point cloud regions based on the motion features, boundary point set, and device semantic category, using these regions as masks for the dynamic targets; 5) Extracting point cloud clusters of dynamic targets from the real-time dense point cloud model using the masks of the dynamic targets.

[0104] Among them, the dense point cloud data obtained from the previous time t-1 and the current time t are based on continuous sampling points. and Timing registration is performed to accurately align spatial positions. Then, the drone's pose data (position, angle, and flight speed) is read, and the "false motion" of the background caused by the drone's flight is calculated. Motion compensation is applied to the static background (stationary equipment such as transformers, steel frames, and busbars) to cancel out the "false motion" of the background, thereby eliminating the interference caused by the drone's own movement. This makes the static background appear completely still in the point cloud, retaining only the real movement of personnel and machinery.

[0105] Among them, motion consistency constraint means that the three-dimensional motion trajectory of the same physical target surface point should have spatial continuity and velocity consistency in continuous time frames. This application compares the point clouds of the two frames after compensation point by point, calculates the position change, motion velocity, motion direction and local motion field distribution of each point, and screens out points whose motion state is completely inconsistent with the surrounding static background, such as position change greater than 0.3m, motion velocity deviation greater than 0.3m / s, motion direction angle difference greater than 15°, and local motion field standard deviation greater than 0.15. All the screened points are clustered into multiple motion significant regions to form an initial dynamic candidate point cloud region set.

[0106] For each candidate point cloud region in the initial dynamic candidate point cloud region set, firstly, the edge refinement and connectivity constraints are performed on each candidate point cloud region by combining the point cloud normal vector, curvature and local geometric structure information. Edge refinement is performed by calculating the rate of change of the point cloud normal vector and the extreme value of curvature to locate the location of abrupt changes in surface geometry. Connectivity constraints are performed by eliminating isolated noise points and break artifacts according to the eight-neighborhood connectivity criterion, while retaining the motion region with complete structure.

[0107] Subsequently, depth mutation detection is performed on the refined candidate point cloud regions, that is, the depth values ​​of the candidate point cloud regions are checked, and the locations where the depth suddenly changes significantly and the surface continuity is interrupted are identified as the boundary point clouds of dynamic targets and formed into a boundary point set.

[0108] Then, a semantic segmentation network is introduced. For each segmented candidate point cloud region, the dense point cloud corresponding to it is input into the semantic segmentation network to extract its corresponding device semantic category. The device semantic category includes static equipment (such as transformers, circuit breakers, steel frames), dynamic targets (such as personnel, construction machinery) and interference items (such as birds, shadows, changes in lighting, metal reflection artifacts).

[0109] Among them, motion features are the aforementioned changes in position, motion speed, motion direction, and local motion field distribution, while semantic features are the semantic categories output by the network.

[0110] The training process of the semantic segmentation network is as follows: a massive labeled industrial scene point cloud dataset is acquired, covering typical operating environments such as substations, wind farms, and transmission lines. The labeled categories strictly correspond to the aforementioned three semantic category labels. The PointNet++ architecture is used as the backbone, and combined with multi-scale feature aggregation and local-global attention mechanisms, the industrial scene point clouds and semantic category labels are trained end-to-end to obtain the trained semantic segmentation network, which can accurately identify the semantic category to which the candidate point cloud region belongs.

[0111] Then, by combining three criteria—motion consistency, depth mutation, and semantic category—it is determined whether the candidate point cloud region is a real dynamic target.

[0112] A true dynamic target requires the candidate point cloud region to simultaneously satisfy the following: significant anomalies in motion features (position change ≥ 0.3m and orientation angle difference > 15°), clear and continuous boundaries of the boundary point set with significant depth abrupt changes (depth gradient magnitude ≥ 0.5), and semantic category belonging to dynamic targets.

[0113] If the above conditions are met simultaneously, the candidate point cloud region can be marked as a real dynamic target; otherwise, it will be removed, thereby solving the misjudgment caused by substation lighting, shadows, metal reflections, and birds.

[0114] Then, the point cloud coordinates of the real dynamic target are used as a mask for the dynamic target to retain only the dynamic target area while occluding the background area. This mask is then used to crop from the real-time dense point cloud model, outputting the point cloud coordinates of the real dynamic target as the point cloud cluster of the dynamic target, i.e.:

[0115]

[0116] in, Indicates time Overall scene point cloud, This indicates the identified background area or the area to be removed. This represents a cluster of point clouds representing the dynamic targets retained after segmentation.

[0117] This application employs dense point cloud real-time reconstruction and dynamic target segmentation technology, enabling accurate extraction of moving targets from complex substation backgrounds. High-density 3D point clouds are generated through multi-view image fusion, and a combination of deep learning and geometric segmentation effectively distinguishes between static backgrounds such as equipment and architecture and dynamic targets such as personnel and vehicles. Thus, even in complex scenarios involving multiple overlapping targets, varying lighting, and partial occlusion, the system can still stably segment the target and assign it a unique ID and trajectory, significantly improving the robustness of target recognition and tracking.

[0118] In some embodiments, based on the point cloud clusters of the dynamic target at each sampling point, cross-frame association and state estimation are performed on the dynamic target to obtain the three-dimensional motion trajectory of the dynamic target. This includes: extracting the three-dimensional geometric features of the point cloud clusters of the dynamic target at each sampling point, and using optical flow to estimate the motion direction and displacement changes of the dynamic target at adjacent sampling points; determining the cross-frame target association relationship of the dynamic target at each sampling point based on the motion direction and displacement changes of the dynamic target, combined with the three-dimensional geometric features; recursively estimating the state of the dynamic target using Kalman filtering based on the cross-frame target association relationship, and fusing the observation results of the current sampling point to determine the optimal estimated position of the dynamic target at the current sampling point; and constructing a continuous three-dimensional trajectory between each sampling point based on the optimal estimated position of the dynamic target to obtain the three-dimensional motion trajectory of the dynamic target.

[0119] Among them, the three-dimensional geometric features serve as cross-frame identity information, including the center position, the bounding box, the local appearance descriptor, and the three-dimensional geometric features. The center position is the centroid coordinate of the point cloud cluster of the dynamic target, the bounding box is the smallest 3D cube that encloses the target, the local appearance descriptor includes the target's texture, color, and surface features, and the three-dimensional geometric features include the shape, normal vector, curvature, and volume of the point cloud.

[0120] Then, the motion direction and displacement change of the dynamic target between adjacent sampling points are calculated using the optical flow method. The optical flow method maps the image domain optical flow to three-dimensional space by analyzing the pixel-level motion vectors between consecutive frames and combining them with depth map constraints, thereby obtaining the true displacement components of the target in the X, Y, and Z directions.

[0121] Then, by utilizing the movement direction and displacement changes of dynamic targets, dynamic targets with continuous movement and consistent appearance at sampling points are identified as the same object. The similarity of their three-dimensional geometric features is compared, and the target association pairs with the highest similarity between adjacent sampling points are selected to establish cross-frame ID mapping, which is the cross-frame target association relationship of dynamic targets at each sampling point, thus avoiding misidentification of different targets.

[0122] Then, when recursively estimating the target state using Kalman filtering, the target's position, velocity, and direction of motion at the previous time t-1 are used as the basis. The predicted position of the target at the current time t is calculated using the Kalman filtering formula, and the actual position of the target at the current time is obtained as the observation result of the current sampling point. The predicted position and the observation result are fused and corrected to determine the optimal estimated position of the dynamic target at the current sampling point, so as to reduce the positioning error caused by detection fluctuations, local occlusion, or noise interference.

[0123] The fusion correction process involves calculating the difference (residual) between the predicted position and the actual observed position detected in the current frame. Based on the magnitude of the residual, weights are assigned using Kalman gain: if the observation is reliable, the correction is more towards the actual position; if the observation is unreliable, the correction is more dependent on the prediction. After weighted fusion, a smoother and more accurate optimal position is obtained, thus completing the correction of the predicted value.

[0124] Furthermore, based on the optimal estimated position of the dynamic target, a continuous, smooth, and physically interpretable three-dimensional trajectory is constructed between each sampling point, represented as:

[0125]

[0126] In the formula, and These represent the target regions extracted at the current time and the previous time, respectively. This represents the cross-frame association and trajectory update process based on optical flow estimation, spatial feature matching, and Kalman filtering. Indicates time The target trajectory.

[0127] Among them, when the target is obscured, missed, or has an incomplete outline in a short period of time, the predicted state of the Kalman filter can still be used to continue generating the trajectory, keeping the trajectory continuous, and restoring the correlation after the target reappears, thus solving the problems of multiple obscurations, easy missed detection, and trajectory breakage in substations.

[0128] This application provides a continuous and reliable three-dimensional target trajectory for ranging by performing stable cross-frame tracking, thereby providing a stable spatial reference to support subsequent high-precision distance calculations and ensuring that the ranging results have millisecond-level response capability and sub-centimeter-level repeatability accuracy in multi-source dynamic occlusion scenarios.

[0129] In some embodiments, the spatial distance between the dynamic target and a preset ranging observation point is calculated in real time based on the three-dimensional motion trajectory of the dynamic target, including: mapping the three-dimensional motion trajectory of the dynamic target and the preset ranging observation point to a unified coordinate system, and discretizing the three-dimensional motion trajectory of the dynamic target in the unified coordinate system to obtain multiple discrete motion points of the dynamic target; for each discrete motion point, calculating the Euclidean distance between the discrete motion point and the preset ranging observation point to determine the spatial distance between the dynamic target and the preset ranging observation point.

[0130] Specifically, the three-dimensional motion trajectory of the dynamic target is mapped to a unified coordinate system with preset ranging observation points to ensure the consistency and comparability of spatial measurements and avoid distance calculation errors caused by coordinate system differences. Based on this, the three-dimensional motion trajectory of the dynamic target is discretized and sampled to obtain multiple discrete motion points of the dynamic target. Then, the Euclidean distance between each discrete motion point and the ranging observation point is calculated independently, thereby obtaining the spatial distance between the target and the preset ranging observation points, i.e.:

[0131]

[0132] In the formula, Indicates time The three-dimensional coordinates of the target center point in a unified coordinate system; Indicates time The three-dimensional coordinates of the distance measurement observation point in a unified coordinate system; It represents the spatial distance between the distance measurement observation point and the target point.

[0133] In some embodiments, the spatial distance between the dynamic target and the preset ranging observation point is calculated in real time based on the three-dimensional motion trajectory of the dynamic target. The method further includes: acquiring ranging error impact data of the current sampling point within the substation; determining the ranging error compensation amount of the current sampling point based on a pre-constructed error impact model and the ranging error impact data; wherein the error impact model is a physical driving model constructed based on the mapping relationship between the ranging error impact data and its corresponding ranging error compensation amount; and correcting the spatial distance according to the ranging error compensation amount to obtain the corrected spatial distance.

[0134] Due to factors such as temperature changes, fluctuations in ambient light, drone vibration, and changes in the long-term operating status of the equipment, the original ranging results may experience zero-point drift, proportional deviation, and accumulation of random errors during continuous operation. To ensure the stability and long-term consistency of the ranging system in complex inspection environments, this application also uses online error compensation to correct the original ranging results in real time.

[0135] Among them, the ranging error-affected data includes environmental data, motion data, equipment data, and historical data. Preferably, the ambient temperature and equipment body temperature are collected by a temperature sensor, the on-site light intensity and exposure status are collected by a light sensor, the vibration amplitude, flight jitter and attitude change of the UAV are collected by an inertial measurement unit, the equipment data includes sensor running time, calibration parameter attenuation value and imaging unit working status, and the historical data includes the measurement residual sequence of past ranging (the deviation between historical measurement values ​​and the true values).

[0136] Then, an error impact model is established, which is obtained by nonlinear polynomial fitting or lightweight residual learning using the offset relationship between the ranging error impact data and the ranging error compensation amount.

[0137] Temperature changes can cause focal length drift in imaging units, changes in sensor internal parameters, or shifts in electronic device response; changes in illumination can lead to unstable image exposure, feature extraction deviations, or fluctuations in depth estimation; equipment vibration and flight jitter can cause image blurring, point cloud registration errors, and target positioning jitter. Based on these influencing factors, the sources of error and their impact under the current measurement conditions are assessed in real time.

[0138] Preferably, during the error compensation process, the ranging error impact data is input into the error impact model, and the corresponding ranging error compensation amount is output. During the error compensation process, the ranging error compensation amount and the spatial distance are fused to obtain the corrected spatial distance, i.e.:

[0139]

[0140] In the formula, This is the distance measurement error compensation amount. This is the corrected spatial distance.

[0141] This application combines multi-view geometric constraints with an online error compensation mechanism to achieve sub-meter accuracy distance measurement. The system not only directly calculates the spatial coordinates and distance of the target based on 3D point clouds, but also establishes an error impact model to compensate in real time for factors such as sensor intrinsic and extrinsic parameter drift and changes in environmental temperature and humidity. This not only makes the ranging results more accurate and reliable, but also significantly reduces the over-reliance on sensor calibration accuracy, maintaining stable measurement performance even during long-term operation.

[0142] In some embodiments, the spatial distance is corrected according to the ranging error compensation amount to obtain the corrected spatial distance. Then, the method further includes: obtaining the reference spatial distance between the dynamic target and the preset ranging observation point; calculating the compensation deviation value based on the corrected spatial distance and the reference spatial distance; when the compensation deviation value exceeds the preset safety threshold, adjusting the parameters of the error influence model with the goal of minimizing the compensation deviation, updating the error influence model, and re-executing the pre-built error influence model based on the updated error influence model, combined with the ranging error influence data, to determine the ranging error compensation amount of the current sampling point, until the compensation deviation value does not exceed the preset safety threshold, and outputting the final corrected spatial distance.

[0143] To maintain ranging accuracy during long-term operation, the system further employs an online calibration strategy based on minimizing errors. Specifically, it obtains an error-free standard distance as a reference spatial distance between the dynamic target and preset ranging observation points, and calculates the difference between the corrected spatial distance and the reference spatial distance to obtain a compensation deviation value. When this deviation exceeds a preset safety threshold, the model needs to be optimized.

[0144] Specifically, with the optimization objective of minimizing the compensation deviation, the parameters (compensation terms and weights) of the error impact model are adjusted and continuously corrected to make the measured value infinitely close to the reference true value. After updating the error impact model, the ranging error compensation amount is recalculated and iteratively corrected until the compensation deviation converges to within the threshold, eliminating the systematic drift caused by long-term operation, temperature changes, and aging. In the complex and harsh environment of the substation, the ranging always maintains centimeter-level high accuracy, long-term stability, and no attenuation.

[0145] The optimization process can be represented as follows:

[0146]

[0147] In the formula, The reference distance can be given by the known distance of the calibrated target, the fixed structural size constraints in the scene, or the results of multi-frame stabilization estimation. This represents the compensation term to be optimized. By continuously minimizing this deviation, the system can gradually suppress systematic drift during long-term operation, improving the stability, repeatability, and environmental adaptability of the ranging results.

[0148] This application utilizes an online calibration model to periodically drive a high-precision calibration target into a common field of view, automatically calibrating and updating the parameters of each sensor. This effectively overcomes parameter drift caused by long-term equipment operation, temperature changes, and mechanical vibrations, maintaining the coordinate system consistency of each sensor and the accuracy of the measurement benchmark without manual intervention, significantly reducing on-site maintenance workload and system downtime.

[0149] After ranging, identification, and calibration are completed, all results are uploaded to the monitoring center, including: 3D trajectory and status information; corrected ranging values; operational status and early warning information; and model update information. This data is transmitted in real time via network communication units, providing a basis for remote diagnostics and status monitoring. The monitoring center can then issue update commands to achieve system-level remote maintenance and strategy optimization.

[0150] Next reference Figure 3 The present application's visual ranging method for substations will be illustrated with a specific embodiment.

[0151] In this specific embodiment, a modular, layered design is adopted in the physical architecture, consisting of a synchronous acquisition unit, an edge processing unit, a control and communication unit, and an online calibration unit. The units are connected via a highly reliable industrial wired network and synchronization signal lines, and data and control flows are transmitted under a unified communication protocol, ensuring system stability and real-time performance.

[0152] The synchronous acquisition unit is deployed at key observation locations within the substation monitoring area to simultaneously acquire multi-dimensional optical information about the scene. The key observation locations can be flexibly determined based on the structure of the work area, target distribution, and field-of-view coverage requirements. For example, in the main equipment area, the synchronous acquisition unit can be installed on poles around the main transformer, frame beams, supports above inspection channels, or fixed monitoring columns to ensure effective coverage of the equipment body, connecting components, and the work area. In the switchyard or busbar area, it can be installed on edge columns, high-level support frames, or fixed supports on the outside of adjacent buildings to obtain a larger overhead viewing angle and reduce equipment obstruction. For scenarios requiring mobile observation, the synchronous acquisition unit can also be integrated into a UAV platform, using an onboard multi-view camera to complete multi-view synchronous acquisition during flight.

[0153] The edge processing unit is preferably installed in a ruggedized edge computing box, monitoring cabinet, or UAV-borne computing platform in the vicinity, and carries core algorithm tasks such as real-time 3D reconstruction, dynamic target segmentation, target tracking, and high-precision ranging. The control and communication unit is responsible for system task scheduling, status management, equipment control, and data interaction with the remote centralized control platform. It can be deployed in the station control cabinet, edge communication box, or background monitoring host.

[0154] The online calibration unit is fixedly installed at one or more common field-of-view locations to periodically perform sensor calibration and correct the system's measurement reference. A common field-of-view location refers to an area that is observable by multiple acquisition units, has a relatively stable spatial position, and is not easily obstructed, such as the intersection of equipment area passageways, the location of fixed poles within the station, a dedicated calibration plate mounting bracket, or a fixed marked area on a building wall. The online calibration unit can use a calibration plate, a characteristic target, reflective markers, or a fixed reference object with known geometric dimensions as the reference target.

[0155] During system operation, the online calibration unit triggers the calibration process according to preset time intervals, equipment start / stop status, or environmental changes. Specifically, the system can perform an initial calibration once daily after power-on and self-test, or perform periodic recalibration after a certain period of continuous operation. Additional calibration can also be triggered when temperature changes exceed thresholds, equipment vibration is abnormal, image clarity decreases, or multi-frame ranging residuals continue to increase. During calibration, the system control acquisition unit synchronously observes the reference target, extracts calibration feature points, calculates the deviation between the current observation results and the reference geometric parameters, and updates the camera's intrinsic and extrinsic parameters, spatial coordinate transformation relationships, and distance compensation parameters accordingly, thereby correcting the system's measurement reference. Through the above-mentioned periodic calibration and dynamic correction mechanism, the system can suppress the accumulation of measurement errors caused by temperature drift, mechanical vibration, installation misalignment, and component aging during long-term operation, maintaining the accuracy and consistency of measurement results.

[0156] 1) Deployment of synchronous acquisition units and signal aggregation:

[0157] This unit typically consists of multiple industrial vision sensors (such as global shutter cameras and depth cameras) and a synchronization controller. The sensors are mounted on robust brackets, and their fields of view must be carefully planned to cover critical work areas, while strictly avoiding strong direct light sources and locations that may cause continuous obstruction, and meeting safety distance requirements. All sensors are connected to the master clock source via synchronization trigger lines to ensure microsecond-level exposure synchronization, guaranteeing the time alignment of multi-view images from the source. Image data output from the sensors is transmitted to the aggregation switch via shielded twisted-pair cables or fiber optic cables. To avoid electromagnetic interference, signal lines and power cables are laid in separate layers, and the shielding layer is grounded at a single point at the access end. The aggregation switch integrates multiple video streams and uploads them to the edge processing unit via high-speed Ethernet.

[0158] 2) Edge processing unit functions and integration:

[0159] This unit serves as the system's "intelligent hub," typically employing a high-performance industrial computer or embedded server. It integrates multiple GPU computing cards and large-capacity memory, responsible for running core algorithms such as dense point cloud reconstruction, dynamic target segmentation and tracking, and online error compensation. It receives raw images and point cloud data from the synchronous acquisition unit and outputs, through parallel computing processes, the target trajectory with 3D coordinates, motion speed, and a unique ID, along with its precise distance measurement information, in real time. This unit possesses powerful local computing and caching capabilities, enabling it to complete all image processing and data analysis within the station, uploading only structured results (such as coordinates, alarm events, and device status), significantly reducing the bandwidth pressure on the backbone network. The unit connects to the switch via standard gigabit / 10-gigabit Ethernet optical or electrical ports. The chassis features a sealed, dustproof design with strong air cooling or liquid cooling active heat dissipation. Internal components are fixed via guide rails and equipped with shock-absorbing devices. The power supply uses dual redundant DC inputs and is equipped with surge protectors and overvoltage isolation modules to adapt to the harsh electromagnetic interference, temperature fluctuations, and humidity changes within the substation environment.

[0160] 3) Task scheduling and interaction of the control and communication units:

[0161] This unit is typically housed in the same ruggedized cabinet as the edge processing unit, directly connected via short-distance fiber optic or network cables to achieve millisecond-level low-latency control. It manages the entire system's operational cycle and logical scheduling: dynamically adjusting the acquisition frame rate according to the work plan, triggering real-time processing tasks, and driving online calibration processes to start / stop on a timed or event-driven basis, while also monitoring the health status of each unit (temperature, load, connection status) in real time. All processing results (target trajectory, alarm logs, system performance indicators) are uniformly encapsulated, compressed, and timestamped within this unit before being uploaded to the remote intelligent auxiliary control platform via the substation's industrial Ethernet or dedicated fiber optic network. The communication protocol supports resume transmission after disconnection and a local ring buffer mechanism to ensure data integrity and prevent data loss. Simultaneously, this unit also receives configuration commands and control strategies from the platform, enabling remote parameter adjustments, algorithm updates, and system start / stop, forming a reliable two-way communication link.

[0162] 4) Automatic calibration mechanism of the online calibration unit:

[0163] This unit is crucial for achieving long-term, high-precision ranging. It consists of a high-precision visual calibration target (such as a checkerboard or concentric circle array), a precision motorized translation stage, and its controller. When the system reaches a preset operating cycle, or when the self-diagnostic module detects excessive sensor parameter drift, the control unit issues a command. The translation stage drives the calibration target to automatically move into the common field of view of all sensors. The system automatically acquires a series of calibration images and calculates and updates the internal parameters (focal length, distortion) and external parameters (relative pose) of each sensor. The new calibration parameters are automatically sent to the edge processing unit to correct the reconstruction and ranging model, thus forming a complete accuracy closed loop from software algorithm to hardware benchmark.

[0164] 5) System power supply and environmental adaptability design:

[0165] The entire system is powered by a unified DC power supply unit within the station. The power lines are equipped with filtering, voltage stabilization, and surge protection units to reduce the impact of power fluctuations and electromagnetic interference on the data acquisition, processing, and communication processes, thereby improving the system's power supply stability and operational reliability. If necessary, independent isolated power supply branches can be set up for critical modules to prevent local anomalies from affecting the overall operation.

[0166] In terms of environmental adaptability, all cabinets, cable trays, interfaces, and exposed connection components meet IP65 or higher protection standards. Key connectors feature anti-detachment and anti-corrosion designs, combined with sealing and reinforcement measures, to improve connection stability under long-term outdoor operating conditions. For outdoor-installed equipment, sun-proof, rain-proof, and vibration-damping structures can also be configured to enhance the system's adaptability to complex environmental factors such as high temperature, humidity, rainfall, and mechanical vibration.

[0167] The equipment layout takes heat dissipation requirements into full consideration. Edge processing units, power modules and communication modules that generate a lot of heat are preferably installed in well-ventilated locations, and can be combined with heat sinks, air cooling or temperature control devices for auxiliary heat dissipation to ensure that the system can continue to operate stably in outdoor high-temperature environments.

[0168] Overall, the system architecture of this application is clearly structured and reliably connected. The four main units—acquisition, processing, control, and calibration—each perform their respective functions while working in concert. Through precise synchronization design, strong anti-interference transmission links, and a periodic automatic calibration mechanism, they jointly ensure that the two core technologies—"dynamic segmentation of dense point clouds" and "online error compensation"—can function stably and reliably in actual substation engineering environments over a long period of time, achieving effective transformation from technical solutions to engineering applications.

[0169] Based on the same inventive concept, this application also provides a substation visual ranging system for implementing the substation visual ranging method described above.

[0170] The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more substation visual ranging system embodiments provided below can be found in the limitations of the substation visual ranging method described above, and will not be repeated here.

[0171] like Figure 4 As shown in the illustration, this application also provides a substation visual ranging system, including:

[0172] The point cloud reconstruction module 100 is used to acquire multi-view image sequences of the target monitoring area in the substation at continuous sampling points, and to perform real-time three-dimensional reconstruction based on the multi-view image sequences to obtain a real-time dense point cloud model for each sampling point.

[0173] The target point cloud segmentation module 200 is used to perform dynamic target segmentation on the scene point cloud of the real-time dense point cloud model for each sampling point, so as to obtain the point cloud cluster of the dynamic target at each sampling point.

[0174] The target tracking module 300 is used to perform cross-frame association and state estimation of the dynamic target based on the point cloud clusters of the dynamic target at each sampling point, so as to obtain the three-dimensional motion trajectory of the dynamic target.

[0175] The target ranging module 400 is used to calculate the spatial distance between the dynamic target and the preset ranging observation point in real time based on the three-dimensional motion trajectory of the dynamic target.

[0176] In some embodiments, the point cloud reconstruction module 100 is used for:

[0177] Multi-view image sequences are obtained by simultaneously acquiring image sequences of the target monitoring area from multiple perspectives using multiple cameras.

[0178] Time synchronization and spatial registration are performed on the multi-view image sequence to obtain the registered multi-view image sequence.

[0179] Feature extraction is performed on the registered multi-view image sequence, and cross-view matching relationships are established based on the extracted features to obtain feature point matching pairs between each view.

[0180] By utilizing multi-view geometric rules and combining feature point matching pairs between different viewpoints, the spatial pose of the camera under each viewpoint is determined.

[0181] Based on the feature point matching pairs between different viewpoints and the camera spatial pose, the three-dimensional coordinates of scene feature points are recovered through triangulation, and the depth information of multi-view image sequences is obtained. The depth information and three-dimensional coordinates of each viewpoint are fused to obtain a dense depth map.

[0182] Incremental registration and fusion of the dense depth maps of each sampling point are performed in a unified coordinate system to generate an initial real-time dense point cloud model for each sampling point.

[0183] Occlusion compensation and surface interpolation optimization are performed on the initial real-time dense point cloud model to obtain the real-time dense point cloud model for each sampling point.

[0184] In some embodiments, the target point cloud segmentation module 200 is used for:

[0185] Temporal registration and background motion compensation are performed on the scene point cloud of the real-time dense point cloud model of each pair of adjacent sampling points to obtain the compensated scene point cloud.

[0186] Based on motion consistency constraints, all motion-salient point cloud regions are extracted from the compensated scene point cloud to construct an initial dynamic candidate point cloud region set.

[0187] For each candidate point cloud region in the initial dynamic candidate point cloud region set, edge refinement and connectivity constraints are applied to the candidate point cloud regions to obtain refined candidate point cloud regions.

[0188] Deep mutation detection is performed on the point cloud within the refined candidate point cloud region to obtain the boundary point set of the refined candidate point cloud region;

[0189] Based on the dense point cloud corresponding to the refined candidate point cloud region, and combined with the preset semantic feature extraction network, the device semantic category of the refined candidate point cloud region is determined.

[0190] The motion features of the refined candidate point cloud regions are obtained. Based on the motion features, boundary point set and device semantic category, candidate point cloud regions of dynamic targets are selected from all refined candidate point cloud regions and used as masks for dynamic targets.

[0191] The point cloud clusters of dynamic targets are extracted from a real-time dense point cloud model using a mask of dynamic targets.

[0192] In some embodiments, the target tracking module 300 is configured to:

[0193] The three-dimensional geometric features of the point cloud clusters of dynamic targets at each sampling point are extracted, and the motion direction and displacement changes of dynamic targets at adjacent sampling points are estimated using the optical flow method.

[0194] Based on the motion direction and displacement changes of the dynamic target, combined with three-dimensional geometric features, the cross-frame target correlation relationship of the dynamic target at each sampling point is determined;

[0195] Based on the cross-frame target correlation, the state of the dynamic target is recursively estimated using Kalman filtering, and the observation results of the current sampling point are fused to determine the optimal estimated position of the dynamic target at the current sampling point;

[0196] Based on the optimal estimated position of the dynamic target, a continuous three-dimensional trajectory between each sampling point is constructed to obtain the three-dimensional motion trajectory of the dynamic target.

[0197] In some embodiments, the target ranging module 400 is used for:

[0198] The three-dimensional motion trajectory of the dynamic target is mapped to a unified coordinate system with preset ranging observation points. Then, the three-dimensional motion trajectory of the dynamic target is discretized and sampled in the unified coordinate system to obtain multiple discrete motion points of the dynamic target.

[0199] For each moving discrete point, calculate the Euclidean distance between the moving discrete point and the preset ranging observation point to determine the spatial distance between the dynamic target and the preset ranging observation point.

[0200] In some embodiments, the system further includes a ranging correction module, configured to:

[0201] Obtain data on the impact of ranging errors at current sampling points within the substation;

[0202] Based on the pre-built error impact model and combined with the ranging error impact data, the ranging error compensation amount for the current sampling point is determined; where the error impact model is a physical driving model built based on the mapping relationship between the ranging error impact data and its corresponding ranging error compensation amount.

[0203] The spatial distance is corrected based on the distance measurement error compensation amount to obtain the corrected spatial distance.

[0204] In some embodiments, the system further includes: a model calibration module, configured to:

[0205] Obtain the reference spatial distance between the dynamic target and the preset ranging observation point, and calculate the compensation deviation value based on the corrected spatial distance and the reference spatial distance;

[0206] When the compensation deviation exceeds the preset safety threshold, the parameters of the error impact model are adjusted with the goal of minimizing the compensation deviation. The error impact model is then updated, and the pre-built error impact model is re-executed based on the updated error impact model. Combined with the ranging error impact data, the ranging error compensation amount of the current sampling point is determined until the compensation deviation does not exceed the preset safety threshold, and the final corrected spatial distance is output.

[0207] like Figure 5 As shown, this application provides an electronic device 10, which includes a memory 20 and a processor 30. The memory 20 stores a computer program. When the computer program is executed by the processor 30, the processor 30 performs the following steps:

[0208] Acquire multi-view image sequences of the target monitoring area within the substation at continuous sampling points, and perform real-time 3D reconstruction based on the multi-view image sequences to obtain a real-time dense point cloud model for each sampling point.

[0209] For each sampling point, dynamic target segmentation is performed on the scene point cloud of the real-time dense point cloud model to obtain the point cloud cluster of dynamic targets at each sampling point;

[0210] Based on the point cloud clusters of dynamic targets at each sampling point, cross-frame association and state estimation are performed on the dynamic targets to obtain the three-dimensional motion trajectory of the dynamic targets.

[0211] Based on the three-dimensional motion trajectory of the dynamic target, the spatial distance between the dynamic target and the preset ranging observation point is calculated in real time.

[0212] In some embodiments, when the computer program is executed by the processor 30, the processor 30 also performs the following steps:

[0213] Multi-view image sequences are obtained by simultaneously acquiring image sequences of the target monitoring area from multiple perspectives using multiple cameras.

[0214] Time synchronization and spatial registration are performed on the multi-view image sequence to obtain the registered multi-view image sequence.

[0215] Feature extraction is performed on the registered multi-view image sequence, and cross-view matching relationships are established based on the extracted features to obtain feature point matching pairs between each view.

[0216] By utilizing multi-view geometric rules and combining feature point matching pairs between different viewpoints, the spatial pose of the camera under each viewpoint is determined.

[0217] Based on the feature point matching pairs between different viewpoints and the camera spatial pose, the three-dimensional coordinates of scene feature points are recovered through triangulation, and the depth information of multi-view image sequences is obtained. The depth information and three-dimensional coordinates of each viewpoint are fused to obtain a dense depth map.

[0218] Incremental registration and fusion of the dense depth maps of each sampling point are performed in a unified coordinate system to generate an initial real-time dense point cloud model for each sampling point.

[0219] Occlusion compensation and surface interpolation optimization are performed on the initial real-time dense point cloud model to obtain the real-time dense point cloud model for each sampling point.

[0220] In some embodiments, when the computer program is executed by the processor 30, the processor 30 also performs the following steps:

[0221] Temporal registration and background motion compensation are performed on the scene point cloud of the real-time dense point cloud model of each pair of adjacent sampling points to obtain the compensated scene point cloud.

[0222] Based on motion consistency constraints, all motion-salient point cloud regions are extracted from the compensated scene point cloud to construct an initial dynamic candidate point cloud region set.

[0223] For each candidate point cloud region in the initial dynamic candidate point cloud region set, edge refinement and connectivity constraints are applied to the candidate point cloud regions to obtain refined candidate point cloud regions.

[0224] Deep mutation detection is performed on the point cloud within the refined candidate point cloud region to obtain the boundary point set of the refined candidate point cloud region;

[0225] Based on the dense point cloud corresponding to the refined candidate point cloud region, and combined with the preset semantic feature extraction network, the device semantic category of the refined candidate point cloud region is determined.

[0226] The motion features of the refined candidate point cloud regions are obtained. Based on the motion features, boundary point set and device semantic category, candidate point cloud regions of dynamic targets are selected from all refined candidate point cloud regions and used as masks for dynamic targets.

[0227] The point cloud clusters of dynamic targets are extracted from a real-time dense point cloud model using a mask of dynamic targets.

[0228] In some embodiments, when the computer program is executed by the processor 30, the processor 30 also performs the following steps:

[0229] The three-dimensional geometric features of the point cloud clusters of dynamic targets at each sampling point are extracted, and the motion direction and displacement changes of dynamic targets at adjacent sampling points are estimated using the optical flow method.

[0230] Based on the motion direction and displacement changes of the dynamic target, combined with three-dimensional geometric features, the cross-frame target correlation relationship of the dynamic target at each sampling point is determined;

[0231] Based on the cross-frame target correlation, the state of the dynamic target is recursively estimated using Kalman filtering, and the observation results of the current sampling point are fused to determine the optimal estimated position of the dynamic target at the current sampling point;

[0232] Based on the optimal estimated position of the dynamic target, a continuous three-dimensional trajectory between each sampling point is constructed to obtain the three-dimensional motion trajectory of the dynamic target.

[0233] In some embodiments, when the computer program is executed by the processor 30, the processor 30 also performs the following steps:

[0234] The three-dimensional motion trajectory of the dynamic target is mapped to a unified coordinate system with preset ranging observation points. Then, the three-dimensional motion trajectory of the dynamic target is discretized and sampled in the unified coordinate system to obtain multiple discrete motion points of the dynamic target.

[0235] For each moving discrete point, calculate the Euclidean distance between the moving discrete point and the preset ranging observation point to determine the spatial distance between the dynamic target and the preset ranging observation point.

[0236] In some embodiments, when the computer program is executed by the processor 30, the processor 30 also performs the following steps:

[0237] Obtain data on the impact of ranging errors at current sampling points within the substation;

[0238] Based on the pre-built error impact model and combined with the ranging error impact data, the ranging error compensation amount for the current sampling point is determined; where the error impact model is a physical driving model built based on the mapping relationship between the ranging error impact data and its corresponding ranging error compensation amount.

[0239] The spatial distance is corrected based on the distance measurement error compensation amount to obtain the corrected spatial distance.

[0240] In some embodiments, when the computer program is executed by the processor 30, the processor 30 also performs the following steps:

[0241] Obtain the reference spatial distance between the dynamic target and the preset ranging observation point, and calculate the compensation deviation value based on the corrected spatial distance and the reference spatial distance;

[0242] When the compensation deviation exceeds the preset safety threshold, the parameters of the error impact model are adjusted with the goal of minimizing the compensation deviation. The error impact model is then updated, and the pre-built error impact model is re-executed based on the updated error impact model. Combined with the ranging error impact data, the ranging error compensation amount of the current sampling point is determined until the compensation deviation does not exceed the preset safety threshold, and the final corrected spatial distance is output.

[0243] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed, performs the following steps:

[0244] Acquire multi-view image sequences of the target monitoring area within the substation at continuous sampling points, and perform real-time 3D reconstruction based on the multi-view image sequences to obtain a real-time dense point cloud model for each sampling point.

[0245] For each sampling point, dynamic target segmentation is performed on the scene point cloud of the real-time dense point cloud model to obtain the point cloud cluster of dynamic targets at each sampling point;

[0246] Based on the point cloud clusters of dynamic targets at each sampling point, cross-frame association and state estimation are performed on the dynamic targets to obtain the three-dimensional motion trajectory of the dynamic targets.

[0247] Based on the three-dimensional motion trajectory of the dynamic target, the spatial distance between the dynamic target and the preset ranging observation point is calculated in real time.

[0248] In some embodiments, when a computer program is executed, it also performs the following steps:

[0249] Multi-view image sequences are obtained by simultaneously acquiring image sequences of the target monitoring area from multiple perspectives using multiple cameras.

[0250] Time synchronization and spatial registration are performed on the multi-view image sequence to obtain the registered multi-view image sequence.

[0251] Feature extraction is performed on the registered multi-view image sequence, and cross-view matching relationships are established based on the extracted features to obtain feature point matching pairs between each view.

[0252] By utilizing multi-view geometric rules and combining feature point matching pairs between different viewpoints, the spatial pose of the camera under each viewpoint is determined.

[0253] Based on the feature point matching pairs between different viewpoints and the camera spatial pose, the three-dimensional coordinates of scene feature points are recovered through triangulation, and the depth information of multi-view image sequences is obtained. The depth information and three-dimensional coordinates of each viewpoint are fused to obtain a dense depth map.

[0254] Incremental registration and fusion of the dense depth maps of each sampling point are performed in a unified coordinate system to generate an initial real-time dense point cloud model for each sampling point.

[0255] Occlusion compensation and surface interpolation optimization are performed on the initial real-time dense point cloud model to obtain the real-time dense point cloud model for each sampling point.

[0256] In some embodiments, when a computer program is executed, it also performs the following steps:

[0257] Temporal registration and background motion compensation are performed on the scene point cloud of the real-time dense point cloud model of each pair of adjacent sampling points to obtain the compensated scene point cloud.

[0258] Based on motion consistency constraints, all motion-salient point cloud regions are extracted from the compensated scene point cloud to construct an initial dynamic candidate point cloud region set.

[0259] For each candidate point cloud region in the initial dynamic candidate point cloud region set, edge refinement and connectivity constraints are applied to the candidate point cloud regions to obtain refined candidate point cloud regions.

[0260] Deep mutation detection is performed on the point cloud within the refined candidate point cloud region to obtain the boundary point set of the refined candidate point cloud region;

[0261] Based on the dense point cloud corresponding to the refined candidate point cloud region, and combined with the preset semantic feature extraction network, the device semantic category of the refined candidate point cloud region is determined.

[0262] The motion features of the refined candidate point cloud regions are obtained. Based on the motion features, boundary point set and device semantic category, candidate point cloud regions of dynamic targets are selected from all refined candidate point cloud regions and used as masks for dynamic targets.

[0263] The point cloud clusters of dynamic targets are extracted from a real-time dense point cloud model using a mask of dynamic targets.

[0264] In some embodiments, when a computer program is executed, it also performs the following steps:

[0265] The three-dimensional geometric features of the point cloud clusters of dynamic targets at each sampling point are extracted, and the motion direction and displacement changes of dynamic targets at adjacent sampling points are estimated using the optical flow method.

[0266] Based on the motion direction and displacement changes of the dynamic target, combined with three-dimensional geometric features, the cross-frame target correlation relationship of the dynamic target at each sampling point is determined;

[0267] Based on the cross-frame target correlation, the state of the dynamic target is recursively estimated using Kalman filtering, and the observation results of the current sampling point are fused to determine the optimal estimated position of the dynamic target at the current sampling point;

[0268] Based on the optimal estimated position of the dynamic target, a continuous three-dimensional trajectory between each sampling point is constructed to obtain the three-dimensional motion trajectory of the dynamic target.

[0269] In some embodiments, when a computer program is executed, it also performs the following steps:

[0270] The three-dimensional motion trajectory of the dynamic target is mapped to a unified coordinate system with preset ranging observation points. Then, the three-dimensional motion trajectory of the dynamic target is discretized and sampled in the unified coordinate system to obtain multiple discrete motion points of the dynamic target.

[0271] For each moving discrete point, calculate the Euclidean distance between the moving discrete point and the preset ranging observation point to determine the spatial distance between the dynamic target and the preset ranging observation point.

[0272] In some embodiments, when a computer program is executed, it also performs the following steps:

[0273] Obtain data on the impact of ranging errors at current sampling points within the substation;

[0274] Based on the pre-built error impact model and combined with the ranging error impact data, the ranging error compensation amount for the current sampling point is determined; where the error impact model is a physical driving model built based on the mapping relationship between the ranging error impact data and its corresponding ranging error compensation amount.

[0275] The spatial distance is corrected based on the distance measurement error compensation amount to obtain the corrected spatial distance.

[0276] In some embodiments, when a computer program is executed, it also performs the following steps:

[0277] Obtain the reference spatial distance between the dynamic target and the preset ranging observation point, and calculate the compensation deviation value based on the corrected spatial distance and the reference spatial distance;

[0278] When the compensation deviation exceeds the preset safety threshold, the parameters of the error impact model are adjusted with the goal of minimizing the compensation deviation. The error impact model is then updated, and the pre-built error impact model is re-executed based on the updated error impact model. Combined with the ranging error impact data, the ranging error compensation amount of the current sampling point is determined until the compensation deviation does not exceed the preset safety threshold, and the final corrected spatial distance is output.

[0279] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, electronic devices, and computer storage media described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0280] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0281] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0282] In the embodiments provided by this invention, it should be understood that the disclosed systems, electronic devices, computer storage media, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0283] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0284] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0285] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods described in the various embodiments of the present invention through a computer device (which may be a personal computer, a server, or a network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0286] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A visual ranging method for substations, characterized in that, include: A multi-view image sequence of the target monitoring area in the substation at continuous sampling points is obtained, and real-time three-dimensional reconstruction is performed based on the multi-view image sequence to obtain a real-time dense point cloud model of each sampling point. For each sampling point, dynamic target segmentation is performed on the scene point cloud of the real-time dense point cloud model to obtain the point cloud cluster of dynamic targets for each sampling point; Based on the point cloud clusters of the dynamic targets at each sampling point, cross-frame association and state estimation are performed on the dynamic targets to obtain the three-dimensional motion trajectory of the dynamic targets. Based on the three-dimensional motion trajectory of the dynamic target, the spatial distance between the dynamic target and the preset ranging observation point is calculated in real time.

2. The substation visual ranging method according to claim 1, characterized in that, The process of acquiring a multi-view image sequence of the target monitoring area within the substation at continuous sampling points, and performing real-time 3D reconstruction based on the multi-view image sequence to obtain a real-time dense point cloud model for each sampling point includes: The multi-view image sequence is obtained by simultaneously acquiring image sequences of the target monitoring area from multiple perspectives using a multi-view camera. The multi-view image sequence is time-synchronized and spatially registered to obtain a registered multi-view image sequence. Feature extraction is performed on the registered multi-view image sequence, and cross-view matching relationships are established based on the extracted features to obtain feature point matching pairs between each view. By utilizing multi-view geometric rules and combining feature point matching pairs between the various viewpoints, the camera spatial pose under each viewpoint is determined. Based on the feature point matching pairs between the various viewpoints and the camera spatial pose, the three-dimensional coordinates of the scene feature points are recovered by triangulation, and the depth information of the multi-view image sequence is obtained. The depth information of each viewpoint is fused with the three-dimensional coordinates to obtain a dense depth map. The dense depth maps of each sampling point are incrementally registered and fused in a unified coordinate system to generate an initial real-time dense point cloud model for each sampling point. Occlusion compensation and surface interpolation optimization are performed on the initial real-time dense point cloud model to obtain the real-time dense point cloud model for each sampling point.

3. The substation visual ranging method according to claim 1, characterized in that, For each of the sampling points, dynamic target segmentation is performed on the scene point cloud of the real-time dense point cloud model to obtain a point cloud cluster of dynamic targets for each sampling point, including: Temporal registration and background motion compensation are performed on the scene point cloud of the real-time dense point cloud model of each pair of adjacent sampling points to obtain the compensated scene point cloud. Based on motion consistency constraints, all motion-salient point cloud regions are extracted from the compensated scene point cloud to construct an initial dynamic candidate point cloud region set. For each candidate point cloud region in the initial dynamic candidate point cloud region set, edge refinement and connectivity constraints are applied to the candidate point cloud regions to obtain refined candidate point cloud regions. Deep mutation detection is performed on the point cloud within the refined candidate point cloud region to obtain the boundary point set of the refined candidate point cloud region; Based on the dense point cloud corresponding to the refined candidate point cloud region, and combined with a preset semantic feature extraction network, the device semantic category of the refined candidate point cloud region is determined. The motion features of the refined candidate point cloud regions are obtained. Based on the motion features, the boundary point set, and the device semantic category, candidate point cloud regions of dynamic targets are selected from all the refined candidate point cloud regions and used as masks for the dynamic targets. The point cloud clusters of the dynamic target are extracted from the real-time dense point cloud model using the mask of the dynamic target.

4. The substation visual ranging method according to claim 1, characterized in that, The step of performing cross-frame association and state estimation of the dynamic target based on the point cloud clusters of the dynamic target at each sampling point to obtain the three-dimensional motion trajectory of the dynamic target includes: The three-dimensional geometric features of the point cloud clusters of dynamic targets at each sampling point are extracted, and the motion direction and displacement changes of dynamic targets at adjacent sampling points are estimated using the optical flow method. Based on the motion direction and displacement changes of the dynamic target, and combined with the three-dimensional geometric features, the cross-frame target association relationship of the dynamic target at each sampling point is determined; Based on the cross-frame target association, the state of the dynamic target is recursively estimated using Kalman filtering, and the observation results of the current sampling point are fused to determine the optimal estimated position of the dynamic target at the current sampling point. Based on the optimal estimated position of the dynamic target, a continuous three-dimensional trajectory between each sampling point is constructed to obtain the three-dimensional motion trajectory of the dynamic target.

5. The substation visual ranging method according to claim 1, characterized in that, The step of calculating the spatial distance between the dynamic target and a preset ranging observation point in real time based on the three-dimensional motion trajectory of the dynamic target includes: The three-dimensional motion trajectory of the dynamic target is mapped to the preset ranging observation point in a unified coordinate system, and the three-dimensional motion trajectory of the dynamic target is discretized and sampled in the unified coordinate system to obtain multiple discrete motion points of the dynamic target. For each of the moving discrete points, the Euclidean distance between the moving discrete point and the preset ranging observation point is calculated to determine the spatial distance between the dynamic target and the preset ranging observation point.

6. The substation visual ranging method according to claim 1, characterized in that, The step of calculating the spatial distance between the dynamic target and a preset ranging observation point in real time based on the three-dimensional motion trajectory of the dynamic target further includes: Obtain the distance measurement error impact data of the current sampling point within the substation; Based on the pre-built error impact model and combined with the ranging error impact data, the ranging error compensation amount for the current sampling point is determined; wherein, the error impact model is a physical driving model built based on the mapping relationship between the ranging error impact data and its corresponding ranging error compensation amount; The spatial distance is corrected based on the distance measurement error compensation amount to obtain the corrected spatial distance.

7. The substation visual ranging method according to claim 6, characterized in that, The step of correcting the spatial distance based on the ranging error compensation amount to obtain the corrected spatial distance further includes: Obtain the reference spatial distance between the dynamic target and the preset ranging observation point, and calculate the compensation deviation value based on the corrected spatial distance and the reference spatial distance; When the compensation deviation value exceeds the preset safety threshold, the parameters of the error influence model are adjusted with the goal of minimizing the compensation deviation, the error influence model is updated, and the pre-built error influence model is re-executed based on the updated error influence model. Combined with the ranging error influence data, the ranging error compensation amount of the current sampling point is determined until the compensation deviation value does not exceed the preset safety threshold, and the final corrected spatial distance is output.

8. A visual ranging system for substations, characterized in that, include: The point cloud reconstruction module is used to acquire multi-view image sequences of the target monitoring area in the substation at continuous sampling points, and to perform real-time three-dimensional reconstruction based on the multi-view image sequences to obtain a real-time dense point cloud model for each sampling point. The target point cloud segmentation module is used to perform dynamic target segmentation on the scene point cloud of the real-time dense point cloud model for each sampling point, so as to obtain the point cloud cluster of the dynamic target for each sampling point; The target tracking module is used to perform cross-frame association and state estimation of the dynamic target based on the point cloud clusters of the dynamic target at each sampling point, so as to obtain the three-dimensional motion trajectory of the dynamic target. The target ranging module is used to calculate the spatial distance between the dynamic target and the preset ranging observation point in real time based on the three-dimensional motion trajectory of the dynamic target.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the substation visual ranging method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the steps of the substation visual ranging method as described in any one of claims 1-7.

Images