Power violation tracing method fusing panoramic image and positioning data

Abstract

This application proposes a method for tracing power violations by integrating panoramic images and positioning data, belonging to the field of traceability technology. The method includes: acquiring the real-time area distortion rate and topological tearing risk factor of the target unit, and jointly generating correction coefficients; generating an effective displacement correction amount based on the correction coefficients and deviation vectors; generating a corrected topological mesh model for each frame based on the effective displacement correction amount; acquiring spatiotemporal behavioral voxel data for all frames based on the corrected topological mesh model; performing a multi-level verification process to generate a voxel data sequence with confidence labels; and performing multi-dimensional feature matching in a preset violation behavior template library to output the violation tracing result. This method can accurately identify violations with different motion trajectories and speed characteristics, significantly reducing false alarm and false negative rates, and realizing a shift from coarse-grained boundary crossing alarms to high-precision, highly robust violation tracing.

Description

Technical Field

[0001] This application relates to the field of traceability technology, and in particular to a method for tracing power violations by fusing panoramic images and location data. Background Technology

[0002] With the increasing demands for operational safety in the power industry, the fusion of panoramic surveillance images and positioning systems for tracing violations has become an important technical means. Existing solutions typically employ panoramic cameras to capture continuous video streams from power work sites, while simultaneously using positioning systems such as UWB, GPS, or BeiDou to acquire the raw location data stream (i.e., three-dimensional physical coordinates) of positioning tags worn by workers in real time. Current technologies attempt to align these two types of data spatiotemporally: first, the positioning data is parsed to obtain the spatial anchor points of the workers; then, perspective projection transformation is used to map these anchor points onto the two-dimensional plane of the panoramic image to aid in identifying the movement trajectories of personnel in specific areas; alternatively, image recognition algorithms are used to extract personnel features, which are then combined with positioning data to determine their absolute coordinates, thereby constructing a human-ground correlation model. This tracing method primarily aims to address the limitations of traditional single-video surveillance, such as limited viewing angles, difficulty in accurately locating the specific physical coordinates of violations, and lack of quantitative analysis of personnel movement states. It seeks to achieve automatic identification, precise location, and post-event tracing of violations.

[0003] However, existing fusion-based source tracing technologies have significant shortcomings in practical applications, mainly manifested in the accumulation of spatial deviations caused by the failure of grid deformation correction in dynamic scenes, and the resulting insufficient behavioral feature recognition capabilities. Specifically, when processing panoramic images, existing technologies often directly construct fixed reference grids or perform simple pixel matching based on the initial static image, ignoring the inherent wide-angle distortion of panoramic lenses and the image geometric deformation caused by equipment vibration and personnel movement. When positioning data is projected onto the image plane, due to the lack of a dynamic correction mechanism based on node topology, once local image deformation occurs, the matching deviation between the positioning anchor point and image features accumulates rapidly, producing a ghosting effect, making it impossible for the positioning coordinates to accurately correspond to the actual violation area. The root cause of this problem is that existing technologies have failed to establish a logic for feeding positioning errors back to image grid correction, and lack a dynamic topology grid update mechanism for adaptively correcting image geometric deformation. This results in the inability to maintain high-precision spatial positioning in complex dynamic scenes, and also an inability to effectively distinguish different violation action patterns such as normal walking, running, falling, or illegal climbing. Ultimately, this leads to low spatial accuracy of the source tracing results, poor robustness of behavior recognition, and a high false alarm rate. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] To achieve the above objectives, this application proposes a method for tracing the source of power violations by fusing panoramic images and location data, including the following steps:

[0006] Step 1: Obtain the panoramic monitoring image sequence and the original location data stream of the positioning system; the panoramic monitoring image sequence includes an initial frame image arranged in chronological order and several subsequent frames.

[0007] Step 2: Construct an initial two-dimensional topological mesh model based on the initial frame image. The initial two-dimensional topological mesh model is composed of several triangular mesh units. Each triangular mesh unit is defined by three virtual mesh nodes, and each virtual mesh node includes planar coordinates.

[0008] Step 3: Based on several frames of subsequent images, select the current frame image to be processed in sequence, parse the original position data stream, extract the three-dimensional physical coordinates of the positioning device under the preset timestamp, and map the three-dimensional physical coordinates to the initial two-dimensional topological mesh model through projection transformation to obtain the two-dimensional projection position of the spatial anchor point;

[0009] Step 4: Based on the spatial anchor point, the two-dimensional projection position is mapped to the initial two-dimensional topological mesh model, and the triangular mesh cell containing the two-dimensional projection position is identified as the target cell; the three virtual mesh nodes of the target cell are respectively used as the first local control node, the second local control node, and the third local control node; the arithmetic mean of the planar coordinates of the first local control node, the second local control node, and the third local control node is used as the geometric center, and the deviation vector is determined according to the geometric center of the target cell and the two-dimensional projection position;

[0010] Step 5: Obtain the real-time area distortion rate of the target unit and the topology tearing risk factors of the first local control node, the second local control node and the third local control node, and generate correction coefficients together. Generate an effective displacement correction amount based on the correction coefficients and the deviation vector, and generate a corrected topology mesh model for each frame based on the effective displacement correction amount.

[0011] Step 6: Based on the corrected topological grid model, acquire the spatiotemporal behavioral voxel data of all frames, execute a multi-level verification process to generate a voxel data sequence with confidence labels, and perform multi-dimensional feature matching in a preset violation behavior template library to output the violation source tracing result.

[0012] Further, the real-time area distortion rate of the target cell and the topology tearing risk factors of the first local control node, the second local control node, and the third local control node are obtained, and correction coefficients are jointly generated. An effective displacement correction is generated based on the correction coefficients and the deviation vector. A corrected topology mesh model for each frame is generated based on the effective displacement correction, including the following steps:

[0013] Step 51: Obtain the historical reference area and the instantaneous area of ​​the target unit in the current frame, and obtain the real-time area distortion rate based on the instantaneous area and the historical reference area;

[0014] Step 52: Obtain the first plane coordinate set of the first local control node, the second local control node and the third local control node in the previous frame, and the second plane coordinate set in the current frame;

[0015] Step 53: Based on the first plane coordinate set and the second plane coordinate set, determine the position vectors of the first local control node, the second local control node, and the third local control node; based on the first plane coordinate set and the second plane coordinate set of the first local control node, the second local control node, and the third local control node, determine the relative scaling ratio of each side length of the target unit;

[0016] Step 54: Calculate the cosine of the angle between adjacent position vectors; determine the topology tear risk factor based on the cosine of the angle and the relative scaling ratio;

[0017] Step 55: If the real-time area distortion rate is greater than the preset area distortion safety threshold, then a correction coefficient is constructed based on the preset area distortion safety threshold and the real-time area distortion rate; if the topology tearing risk factor is greater than the preset tearing risk safety threshold, then the correction coefficient is defined as 0; otherwise, the preset base coefficient is used as the correction coefficient.

[0018] Step 56: Generate an effective displacement correction amount based on the correction coefficient and the deviation vector, and correct the first local control node, the second local control node, the third local control node and the virtual mesh node based on the effective displacement correction amount to obtain a corrected topological mesh model.

[0019] Further, based on the planar coordinates of the first local control node, the second local control node, and the third local control node and the effective displacement correction amount, the first optimized node coordinates are obtained; based on the virtual mesh nodes inside the target unit, the first effective displacement correction amount is constructed through the effective displacement correction amounts of the first local control node, the second local control node, and the third local control node, and the first effective displacement correction amount is superimposed on the planar coordinates based on the virtual mesh nodes inside the target unit to obtain the second optimized node coordinates.

[0020] Further, based on the virtual mesh nodes of the triangular mesh units outside the target unit, search for the N closest corrected nodes, where the corrected nodes are the first local control node, the second local control node, the third local control node, and the virtual mesh node of the target unit; determine the first distance between the virtual mesh nodes of the triangular mesh units outside the target unit and the corrected nodes, determine the normalized weight coefficient based on the first distance, and obtain the coordinates of the third optimized node according to the effective displacement correction amount of the corrected node and the normalized weight coefficient.

[0021] Furthermore, based on the first optimized node coordinates, the second optimized node coordinates, and the third optimized node coordinates, the corrected topology mesh model is obtained.

[0022] Furthermore, a multi-level verification process is executed to generate a voxel data sequence with confidence labels, and multi-dimensional feature matching is performed in a pre-defined violation behavior template library to output the violation tracing results, including the following steps:

[0023] Step 61: Based on the spatiotemporal behavior voxel data, obtain the texture gradient variance value and the signal-to-noise ratio of the positioning signal of the triangular mesh cell where the target character is located; if the texture gradient variance value is less than the first threshold or the signal-to-noise ratio of the positioning signal is less than the second threshold, it is determined that there is environmental interference in the current spatiotemporal behavior voxel data, and the spatiotemporal behavior voxel data is processed by compensation branch: the spatiotemporal behavior voxel data of adjacent frames are weighted and smoothed to generate the first voxel data and mark it as the second confidence state; otherwise, the spatiotemporal behavior voxel data is transmitted to step 62 to execute the first level judgment process;

[0024] Step 62: Determine the instantaneous acceleration vector based on the position vector, and obtain the instantaneous acceleration rate of change based on the instantaneous acceleration vector; obtain the standard instantaneous acceleration rate of change sequence based on the preset violation behavior template library, calculate the Euclidean distance between the instantaneous acceleration rate of change and each standard instantaneous acceleration rate of change in the standard instantaneous acceleration rate of change sequence, and normalize the Euclidean distance to obtain the dynamic matching score;

[0025] Step 63: If the dynamic matching score is greater than the third threshold, then mark the spatiotemporal behavior voxel data as a first-level confidence state; otherwise, proceed to step 64.

[0026] Step 64: Obtain the planar coordinate sequence of the virtual grid node where the target character is located in five consecutive frames, and construct a discrete path point set; select three consecutive points in the path point set, calculate the absolute value of the difference between the slope of the line connecting the first two points and the slope of the line connecting the last two points, and use it as the local curvature value; obtain the normalized path curvature index based on the local curvature value; if the normalized path curvature index is greater than the fourth threshold, then mark the spatiotemporal behavior voxel data as a second-level confidence state; otherwise, then remove the spatiotemporal behavior voxel data.

[0027] Step 65: Integrate the first voxel data, the spatiotemporal behavioral voxel data, and the corresponding confidence labels into a voxel data sequence;

[0028] Step 66: Generate the source tracing result based on the confidence labels of the voxel data sequence.

[0029] Compared with existing technologies, the power violation tracing method integrating panoramic images and positioning data provided in this application completely solves the problems of positioning drift and missing behavioral features caused by grid deformation in dynamic scenes by existing technologies through identification, quantification, correction, and iteration. Specifically, this application first accurately identifies the target unit in the initial two-dimensional topological grid model and simultaneously obtains the real-time area distortion rate of the unit and the topological tear risk factors of the first, second, and third local control nodes, and quantifies the degree of geometric deformation and the stability of node connections in multiple dimensions; based on the real-time area distortion rate and the topological tear risk factors, a correction coefficient is generated, and the effective displacement correction amount is calculated by combining the deviation vector, and the grid node position is fine-tuned, thereby dynamically generating the corrected topological grid model of the current frame from the state of the previous frame, realizing the frame-by-frame iterative update of the grid model as the field environment changes, and eliminating the spatial anchor point misalignment caused by the failure of the static model from a physical mechanism. Building upon this foundation, this application further utilizes the generated corrected topological mesh to construct spatiotemporal behavioral voxel data and introduces a multi-dimensional quality assessment system. This system not only considers texture gradient variance to filter visual noise but also combines the signal-to-noise ratio of the positioning signal to ensure data source reliability. Simultaneously, it deeply mines dynamic features such as instantaneous acceleration vectors, instantaneous acceleration change rates, and normalized path curvature indices, calculating dynamic matching scores to accurately distinguish complex movement patterns such as walking, running, falling, or climbing. Finally, based on the aforementioned comprehensive assessment results, the spatiotemporal behavioral voxel data is intelligently filtered to generate high-confidence labels and output source tracing results. This scheme, by tightly integrating the frame-by-frame self-correction mechanism of the dynamic mesh with the refined extraction of high-order dynamic features, ensures that even under strong interference environments such as equipment vibration and rapid personnel movement, the positioning coordinates are always accurately mapped to the real physical area. Furthermore, it can accurately identify violations with different movement trajectories and speed characteristics, significantly reducing false alarm and false negative rates, and realizing a transformation from coarse-grained boundary crossing alarms to high-precision, highly robust violation source tracing. Attached Figure Description

[0030] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0031] Figure 1 A flowchart illustrating the method for tracing power violations by fusing panoramic images and location data, provided in this application embodiment. Detailed Implementation

[0032] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0033] The following describes a method for tracing power violations by fusing panoramic images and location data, according to an embodiment of this application, with reference to the accompanying drawings.

[0034] like Figure 1 As shown, the method for tracing the source of power violations by integrating panoramic images and location data includes the following steps:

[0035] Step 1: Obtain the panoramic monitoring image sequence and the original location data stream of the positioning system; the panoramic monitoring image sequence includes an initial frame image arranged in chronological order and several subsequent frames.

[0036] This embodiment deploys a monitoring subsystem containing a wide-angle panoramic camera array. This subsystem continuously acquires real-time video streams from the power operation site at a fixed frame rate (e.g., 30fps or 60fps), forming a panoramic monitoring image sequence strictly ordered by timestamps (t0 is the initial frame image, t1, t2, ..., ty are several subsequent frames, and y is greater than 2). The initial frame image is marked as the reference frame at the system startup time or the time of manual triggering, and is used to establish the initial two-dimensional topological mesh model and global coordinate system reference. The subsequent frames are used as dynamic observation data to capture changes in the position and posture of the workers and minor disturbances in the field environment. Simultaneously, a high-precision positioning system (such as a base station-tag network based on UWB ultra-wideband technology or a BeiDou / GPS dual-mode positioning terminal) is used to collect the raw location data stream of the positioning tags worn by the workers in real time. This data stream not only contains three-dimensional physical coordinates, but also includes high-precision timestamps, signal strength indicators (RSSI), and positioning signal-to-noise ratio (SNR) and other metadata. Furthermore, the sampling frequency of the positioning data must be strictly synchronized with the acquisition frequency of the panoramic images (e.g., microsecond-level alignment via the PTP protocol) to ensure that each frame corresponds to a precise spatial anchor point at the same moment. During data acquisition, the panoramic image sequence undergoes preliminary processing including dehazing, exposure compensation, and distortion correction to eliminate the effects of inherent optical distortion caused by the lens. The raw positioning data stream is then filtered and smoothed to remove abnormal jump points caused by multipath effects or signal obstruction.

[0037] Step 2: Construct an initial two-dimensional topological mesh model based on the initial frame image. The initial two-dimensional topological mesh model is composed of several triangular mesh units. Each triangular mesh unit is defined by three virtual mesh nodes, and each virtual mesh node includes planar coordinates.

[0038] In one specific embodiment of this application, the process of constructing an initial two-dimensional topological mesh model based on the initial frame image aims to establish a digital spatial reference that strictly corresponds to the physical environment of the power operation site. Specifically, the acquired initial frame image is first subjected to feature point extraction and semantic segmentation to identify key fixed facilities within the substation (such as tower bases, transformer casings, insulator strings, fences, etc.) and ground texture boundaries. Subsequently, using these identified feature points as anchor points, an initial two-dimensional topological mesh model covering the entire effective monitoring area is generated in the image plane coordinate system using a triangulation algorithm. This model consists of a large number of tightly joined triangular mesh units. Each triangular mesh unit is defined by three virtual mesh nodes, and each virtual mesh node is assigned a unique identifier and precise planar coordinates (x, y). These coordinates not only correspond to the image pixel positions but are also mapped to real physical world coordinates (e.g., latitude and longitude in meters or relative coordinates) through a pre-calibrated intrinsic and extrinsic parameter matrix, thereby establishing a one-to-one correspondence between pixels and physical coordinates. During the construction process, for linear equipment (such as crossarms and conductors), the edges of adjacent triangles are forcibly constrained to be parallel to the equipment to ensure that the mesh edges closely follow the equipment outline. For circular or curved equipment (such as insulator discs and cylindrical towers), an adaptive densification strategy is adopted, increasing the density of virtual mesh nodes in areas with large curvature changes to improve the fidelity of local details. In addition, each virtual mesh node is associated with initial state attributes, including its mesh cell ID, preset physical attribute labels (such as conductor, insulator, safe zone, or dangerous zone), and an initial topology connection table. This table records the adjacency information of each node with other nodes, forming a complete graph data structure.

[0039] Step 3: Based on several subsequent images, select the current frame image to be processed sequentially, parse the original position data stream, extract the three-dimensional physical coordinates of the positioning device under the preset timestamp, and map the three-dimensional physical coordinates to the initial two-dimensional topological mesh model through projection transformation to obtain the two-dimensional projection position of the spatial anchor point.

[0040] In this embodiment, the current frame image to be processed is selected sequentially from the panoramic monitoring image sequence according to time order, and the raw position data stream reading interface of the positioning system is triggered. During the reading process, according to the preset timestamp alignment strategy (such as high-precision clock synchronization or interpolation algorithm based on PTP protocol), the three-dimensional physical coordinates that perfectly match the acquisition time of the current frame image are accurately extracted from the continuous positioning data stream. These coordinates represent the spatial position of the positioning device (such as the UWB tag worn by the operator) in the real world. Subsequently, using the pre-calibrated camera intrinsic parameter matrix (including focal length, principal point coordinates, and distortion coefficients) and extrinsic parameter matrix (describing the rotation and translation relationship between the camera coordinate system and the world coordinate system), a strict perspective projection transformation is performed: the extracted three-dimensional physical coordinates are transformed to the camera normalized plane, and then further mapped to the pixel coordinate system of the initial two-dimensional topological mesh model to obtain the corresponding two-dimensional projection position.

[0041] Step 4: Based on the spatial anchor point, the two-dimensional projection position is mapped to the initial two-dimensional topological mesh model, and the triangular mesh cell containing the two-dimensional projection position is identified as the target cell; the three virtual mesh nodes of the target cell are respectively used as the first local control node, the second local control node, and the third local control node; the arithmetic mean of the planar coordinates of the first local control node, the second local control node, and the third local control node is used as the geometric center, and the deviation vector is determined according to the geometric center of the target cell and the two-dimensional projection position.

[0042] First, using the 2D projection position generated in the previous step as the query key, the graph data structure of the initial 2D topological mesh model is traversed. A point-to-surface inclusion algorithm (such as ray casting or centroid coordinate method) is used to quickly retrieve and locate the unique triangular mesh cell containing the 2D projection position, marking it as the target cell at the current moment. Then, all three virtual mesh nodes defining the boundary are extracted from the topological connectivity table of this target cell, and assigned as the first local control node, second local control node, and third local control node according to a preset node index order. These three nodes constitute the smallest topological primitive describing the geometry of the current region. Next, the planar coordinates of these three local control nodes stored in the initial model are read, and their arithmetic mean is calculated to determine the geometric center of the target cell. This center represents the theoretical centroid position of the mesh cell in a distortion-free state. Based on this, a vector is constructed pointing from the geometric center to the 2D projection position of the spatial anchor point, i.e., the deviation vector: obtained by calculating the difference between the 2D projection position and the geometric center of the target cell. The magnitude of this vector represents the offset distance of the actual target position relative to the theoretical center of the mesh, while its direction angle indicates the specific orientation of the offset. The generation of this deviation vector not only reveals the relative distribution characteristics of the target points within the grid, but also provides a crucial geometric reference for calculating the real-time area distortion rate in subsequent steps. When the deviation vector is too large or its direction is abnormal, it often indicates that the grid in that area may have undergone unexpected stretching, compression, or rotational deformation, thereby triggering an in-depth assessment of the risk factors of topological tearing of local control nodes. This ensures that the subtle changes in the grid topology caused by rapid personnel movement or equipment vibration can be keenly captured, achieving a leap from static coordinate matching to dynamic deformation perception.

[0043] Step 5: Obtain the real-time area distortion rate of the target unit and the topology tearing risk factors of the first local control node, the second local control node and the third local control node, and generate correction coefficients together. Generate an effective displacement correction amount based on the correction coefficients and the deviation vector, and generate a corrected topology mesh model for each frame based on the effective displacement correction amount.

[0044] Step 51: Obtain the historical reference area and the instantaneous area of ​​the target unit in the current frame, and obtain the real-time area distortion rate based on the instantaneous area and the historical reference area.

[0045] This embodiment calculates the displacement rate of the device between adjacent time stamps in real time (obtained by dividing the distance between the three-dimensional physical coordinates of adjacent time stamps by the time interval between adjacent time stamps) and the signal-to-noise ratio of the positioning signal. When the displacement rate exceeds a preset threshold for violent motion (e.g., 0.8 m / s) or the signal-to-noise ratio of the positioning signal falls below a low threshold, causing discontinuous jumps in distance measurement (e.g., a distance jump exceeding 1.5 m within a single frame), it is determined that there is a risk of violent motion or signal jump in the current scene; conversely, if the above indicators are all in a stable range, it is determined to be a static stable state. For the static stable state, this embodiment assumes that the local topology of the target unit is not destroyed, and the planar coordinates of the three virtual mesh nodes (first, second, and third local control nodes) remain as the historical reference coordinates at the time of initial construction (i.e., the planar coordinates of the three virtual mesh nodes in the initial two-dimensional topology model). Using this set of static coordinates, the instantaneous area is calculated using the vector cross product formula. This instantaneous area is also the historical reference area, and the real-time area distortion rate is: |instantaneous area - historical reference area| / historical reference area. The resulting real-time area distortion rate approaches zero, indicating that the mesh is in a healthy state.

[0046] When the displacement rate exceeds a preset threshold for violent motion or a signal jump occurs, the target unit is determined to have entered a dynamic deformation state. The two-dimensional projection position of the target unit detected in the current frame is defined as the geometric centroid constraint point of the current mesh after deformation, serving as the projection center of the deformed shape on the imaging plane. Using the camera's perspective projection matrix and the current three-dimensional physical height information of the positioning device, combined with optimization algorithms such as the least squares method or the nonlinear Gauss-Newton iteration method, a new set of instantaneous plane coordinates is solved in reverse. The core mathematical logic is as follows: Suppose the target mesh consists of three virtual nodes p1, p2, and p3, and the instantaneous three-dimensional coordinates of the i-th local control node to be determined are (x... i y i , z i ), z a With the physical height information of the positioning device fixed or constrained to a known value, the system constructs an objective function aimed at finding a new set of node coordinates. This ensures that the system simultaneously satisfies two key constraints: first, the projection coincidence constraint, which minimizes the Euclidean distance between the centroid of the triangle formed by the new node after perspective projection transformation and the 2D projection position detected in the current frame; and second, the physical continuity constraint, which ensures that the variation in the side length of the triangle formed by the new node does not exceed the material limit. , Let be the coordinates of the i-th local control node. Let J be the coordinates of the j-th local control node. Let i be the historical reference coordinates of the i-th local control node. Let J be the historical reference coordinates of the j-th local control node. The maximum allowable deformation threshold is defined by introducing the aforementioned dual constraints. This transforms the originally underdetermined single-point back-calculation problem into an optimal solution problem with a regularization term, thereby calculating the optimal solution that simultaneously satisfies the alignment of the projection centroid and the physical continuity of the side lengths. This yields the instantaneous planar coordinates that reflect the true deformation state. Using these newly obtained instantaneous planar coordinates, the instantaneous area and real-time area distortion rate of the target element are recalculated using the vector cross product formula.

[0047] Step 52: Obtain the first plane coordinate set of the first local control node, the second local control node and the third local control node in the previous frame, and the second plane coordinate set in the current frame.

[0048] Step 53: Based on the first set of planar coordinates and the second set of planar coordinates, determine the position vectors of the first local control node, the second local control node and the third local control node; based on the first set of planar coordinates and the second set of planar coordinates of the first local control node, the second local control node and the third local control node, determine the relative scaling ratio of each side length of the target unit.

[0049] In one specific embodiment of this application, in order to quantify the deformation risk of the target unit at the geometric topology level, the system performs a series of refined vector operations and ratio calculations based on the first plane coordinate set of the previous frame and the second plane coordinate set of the current frame. Specifically, the two coordinate sets are first parsed, and the i-th local control node constituting the target unit (i.e., the first, second, and third local control nodes) is extracted from them. Time and The planar position vector at time t is denoted as and , For each local control node, its position vector is calculated using vector subtraction: , This represents the position vector of the i-th local control node, which intuitively reflects the direction and distance of movement of each node within the time span.

[0050] Calculate the relative scaling ratio of the three sides of the target cell between two consecutive frames. For each pair of connected nodes (i,j), first calculate its scaling ratio in the previous frame. Original side length of time ,calculate Instantaneous side length of time Finally, the ratio of the original side length to the instantaneous side length is used as the relative scaling ratio. Represents the j-th local control node in The planar position vector at time t. Represents the j-th local control node in The planar position vector at time t. This ratio directly characterizes the degree of stretching (greater than 1) or compression (less than 1) of the mesh edges, and is the most direct evidence for detecting physical deformation of the mesh.

[0051] Step 54: Calculate the cosine value of the angle between adjacent position vectors; determine the topology tear risk factor based on the cosine value of the angle and the relative scaling ratio.

[0052] Select any two distinct nodes (e.g., node i and node j), and calculate the cosine of the angle between their respective position vectors. , Let represent the cosine of the angle between the i-th local control node and the j-th local control node. This represents the position vector of the j-th local control node. This index is used to determine the relative motion trend of adjacent nodes: if A value close to 1 indicates that the two nodes move in almost the same direction and at nearly the same speed, and the mesh maintains a rigid structure; if... A significant decrease or even a negative value suggests that relative torsion, shearing, or back separation has occurred between nodes, indicating a severe precursor to mesh tearing.

[0053] , Represents the risk factor for topological tearing. This represents the relative scaling ratio between the i-th local control node and the j-th local control node. Represents the first weighting ratio. This represents the second weight ratio, min represents the operation of taking the minimum value, and max represents the operation of taking the maximum value.

[0054] It captures the degree of dispersion of the motion direction between nodes (i.e., shear risk). It captures the magnitude of maximum unilateral stretching or compression (i.e., fracture risk). It can not only determine whether the mesh has deformed, but also further distinguish whether this deformation tends to be overall twisting or localized breakage.

[0055] Step 55: If the real-time area distortion rate is greater than the preset area distortion safety threshold, then a correction coefficient is constructed based on the preset area distortion safety threshold and the real-time area distortion rate; if the topology tearing risk factor is greater than the preset tearing risk safety threshold, then the correction coefficient is defined as 0; otherwise, the preset base coefficient is used as the correction coefficient.

[0056] In this embodiment, the area distortion safety threshold is the maximum allowable relative change rate of the mesh before significant deformation occurs. It is usually determined based on the elastic limit of the material or the geometric accuracy requirements of the system. In this embodiment, it is set to 0.15 (i.e., 15%). The tear risk safety threshold is the critical value for determining the risk of topological fracture of the mesh structure. It is usually calculated based on the experimentally measured maximum relative motion angle and stretch ratio between nodes. In this embodiment, it is set to 0.8.

[0057] If the real-time area distortion rate (e.g., 0.2) is greater than the preset area distortion safety threshold, it indicates that the mesh has been overstretched or compressed. A correction coefficient is constructed based on the preset area attenuation coefficient k (0.5): G0×(1-k(m-T1)), where G0 is the preset base coefficient, preferably 1, m is the real-time area distortion rate, and T1 is the preset area distortion safety threshold.

[0058] If the topology tearing risk factor (e.g., 0.9) is greater than the preset tearing risk safety threshold, it indicates that the internal structure of the mesh is in an extremely high-risk state and there is a possibility of instantaneous breakage. The correction coefficient is then forcibly defined as 0.

[0059] Conversely, if none of the above conditions are met, the base coefficient is kept at 1 as the correction coefficient.

[0060] Step 56: Generate an effective displacement correction amount based on the correction coefficient and the deviation vector, and correct the first local control node, the second local control node, the third local control node and the virtual mesh node based on the effective displacement correction amount to obtain a corrected topological mesh model.

[0061] In this embodiment, the effective displacement correction is obtained by multiplying the correction coefficient by the deviation vector. Based on the Euclidean distance calculation formula, the connection distance between the two-dimensional projection position and the first local control node, the second local control node, and the third local control node is calculated. The reciprocal of the connection distance is used as the allocation weight for each local control node. The independent displacement correction of the first local control node, the second local control node, and the third local control node is obtained by multiplying the effective displacement correction by the corresponding node allocation weight.

[0062] The coordinates of the first optimized node are obtained based on the planar coordinates of the first local control node, the second local control node, and the third local control node and the effective displacement correction amount.

[0063] The first optimized node coordinates are obtained by superimposing the planar coordinates of the first local control node, the second local control node, and the third local control node with their corresponding independent displacement corrections.

[0064] Based on the virtual mesh nodes inside the target unit, a first effective displacement correction is constructed using the effective displacement corrections of the first local control node, the second local control node, and the third local control node. The first effective displacement correction is then superimposed onto the planar coordinates of the virtual mesh nodes inside the target unit to obtain the second optimized node coordinates.

[0065] First, the centroid coordinates of the target element's triangle are calculated. The three components of these centroid coordinates are then used as weighting factors for three independent displacement corrections. A weighted sum of these weighting factors and the independent displacement corrections is then performed to obtain the first effective displacement correction. This first effective displacement correction is then superimposed onto the planar coordinates of the virtual mesh nodes within the target element to obtain the second optimized node coordinates.

[0066] Based on the virtual mesh nodes of the triangular mesh units outside the target unit, search for the N nearest corrected nodes, where N is greater than or equal to 2. The corrected nodes are the first local control node, the second local control node, the third local control node, and the virtual mesh node of the target unit. Determine the first distance between the virtual mesh nodes of the triangular mesh units outside the target unit and the corrected nodes. Determine the normalized weight coefficient based on the first distance. Obtain the coordinates of the third optimized node based on the effective displacement correction amount of the corrected nodes and the normalized weight coefficient.

[0067] The first distance between the virtual mesh nodes of the triangular mesh elements outside the target element and the first local control node, the second local control node, the third local control node and the virtual mesh nodes inside the target element is calculated using the Euclidean distance formula. The sum of all the first distances is used as the denominator, and the current first distance is used as the numerator to obtain the normalized weight coefficient. The effective displacement correction of the corrected node is multiplied by the normalized weight coefficient to obtain the coordinates of the third optimized node.

[0068] Based on the coordinates of the first optimized node, the second optimized node, and the third optimized node, the corrected topology mesh model is obtained.

[0069] After the above processing, the coordinates of all virtual grid nodes in the initial two-dimensional topological mesh model are corrected, thus obtaining the corrected topological mesh model.

[0070] If there are still unprocessed subsequent frame images, use the corrected topological mesh model of the current frame as the reference model for the next frame image, and return to step 3 to perform iterative model updates until the corrected topological mesh models of all frames are obtained.

[0071] Step 6: Based on the corrected topological grid model, acquire the spatiotemporal behavioral voxel data of all frames, execute a multi-level verification process to generate a voxel data sequence with confidence labels, and perform multi-dimensional feature matching in a preset violation behavior template library to output the violation source tracing result.

[0072] In three-dimensional space, the target monitoring area (ROI) is precisely defined, and discretized into a series of regularly arranged cubic units according to a preset spatial resolution. These cubic units constitute the spatiotemporal behavioral voxels. Each voxel in this three-dimensional space is traversed, and local mesh fragments covering the voxel's range are accurately located using ray casting algorithms or spatial interpolation techniques. For each located local mesh fragment, the following dual feature extraction operations are performed: First, by extracting the image texture information covering the mesh surface, its Sobel gradient magnitude is calculated to quantify the intensity and directionality of texture changes in the region in space, thereby characterizing visual features such as micro-cracks, wear, or surface roughness on the material surface; second, the position vectors of all control nodes located within the voxel in the corrected topology mesh relative to the initial time (i.e., the difference between the current frame position vector and the initial time position vector) are obtained to characterize the dynamic deformation state of the region. After feature extraction of a single frame, the aforementioned multi-dimensional information is fused and encoded to generate a standardized voxel data block containing spatial coordinates (x, y, z), a time series stamp, the mean Sobel gradient magnitude (reflecting static texture features), and the average position vector (reflecting dynamic deformation features). This data processing method based on 3D voxelization not only achieves deep fusion of 2D mesh models and 3D positioning data, but also, by introducing joint encoding of Sobel gradients and position vectors, makes each voxel an independent data unit containing rich spatiotemporal information.

[0073] Step 61: Based on the spatiotemporal behavior voxel data, obtain the texture gradient variance value and the signal-to-noise ratio of the positioning signal of the triangular mesh cell where the target character is located; if the texture gradient variance value is less than the first threshold or the signal-to-noise ratio of the positioning signal is less than the second threshold, it is determined that there is environmental interference in the current spatiotemporal behavior voxel data, and the spatiotemporal behavior voxel data is processed by compensation branch: the spatiotemporal behavior voxel data of adjacent frames are weighted and smoothed to generate the first voxel data and mark it as the second confidence state; otherwise, the spatiotemporal behavior voxel data is transmitted to step 62 to execute the first level judgment process.

[0074] In one specific embodiment of this application, to achieve high-precision tracking of the target person and suppression of environmental interference, a strong correlation mapping mechanism between spatiotemporal behavioral voxel data and the target person is first established. The spatiotemporal behavioral voxel data serves as a holographic digital representation of the scene, where each voxel unit stores the geometric topological information and texture features at that location. When the target person enters the monitoring field of view, the physical space occupied by their body corresponds to a specific subset region in the spatiotemporal voxel grid. The system uses a pre-trained target detection algorithm to output the target person's 3D bounding box and key component coordinates in real time, thereby accurately locating and extracting the target voxel clusters belonging to that person.

[0075] Based on this correlation, a dual quality assessment is performed on the target voxel cluster: first, the texture gradient variance of all triangular mesh units within the cluster is calculated based on the Sobel gradient magnitude of the spatiotemporal behavioral voxels, used to measure the clarity and richness of the surface texture of the person (such as clothing wrinkles and facial features); second, the signal-to-noise ratio (SNR) of the positioning signal corresponding to the cluster is extracted to evaluate the reliability of the measurement of the person's spatial coordinates at the current moment. Regarding the threshold setting criteria, the first threshold is set based on the statistical distribution of the texture features of the person under normal lighting, taken as twice the standard deviation of the normal texture background variance; in this embodiment, it is set to 0.045. If this value is less than 0.045, it indicates that the target person is overexposed, severely motion-blurred, or obscured by a solid-color background, resulting in missing texture features; the second threshold is set based on the sensor's minimum resolvable distance error under low illumination or multipath effects; in this embodiment, it is set to 6.5 dB. If the positioning signal SNR is lower than 6.5 dB, it indicates that the positioning signal of the target person is severely interfered with by environmental noise (such as radar clutter and visual noise), reducing the reliability of the coordinate data.

[0076] If the voxel cluster corresponding to the target person is detected to meet any of the above conditions, it is determined that the spatiotemporal behavioral voxel data of the target person has significant environmental interference, and a compensation branch is triggered. In this branch, the spatiotemporal behavioral voxel data corresponding to the target position in adjacent frames are adaptively weighted and smoothed using the continuity of the target person's motion in the spatiotemporal domain. The weight allocation follows the principle of temporal proximity (e.g., 0.3 for the current frame, 0.35 for the previous frame, and 0.35 for the next frame), generating the first voxel data after denoising and repair. This data is marked as a second-level confidence state, indicating that although it is interfered with, it can still be used for subsequent analysis after correction. Conversely, if the target person's texture features are clear and the positioning signal is reliable, the system will directly transmit the original spatiotemporal behavioral voxel data to step 62, entering the first-level judgment process for more refined behavior recognition or anomaly detection. Through this dynamic quality assessment and hierarchical processing strategy based on the spatiotemporal voxel cluster of the target person, the problem of lost or misjudged tracking of people in complex environments is effectively solved, ensuring the robustness of the human-computer interaction and monitoring system.

[0077] Step 62: Determine the instantaneous acceleration vector based on the position vector, and obtain the instantaneous acceleration rate of change based on the instantaneous acceleration vector; obtain the standard instantaneous acceleration rate of change sequence based on the preset violation behavior template library, calculate the Euclidean distance between the instantaneous acceleration rate of change and each standard instantaneous acceleration rate of change in the standard instantaneous acceleration rate of change sequence, and normalize the Euclidean distance to obtain the dynamic matching score.

[0078] Based on the position vector, the instantaneous acceleration vector of the worker's limbs or key equipment is continuously calculated in the time dimension using a numerical differentiation algorithm. The instantaneous acceleration vector is then differentiated again in time to accurately calculate the instantaneous acceleration change rate, which characterizes the sudden change in the intensity of the motion. This indicator can effectively capture typical dangerous moments in power operations, such as: violent operation when accidentally opening or closing a disconnecting switch, sudden instability when illegally climbing a tower, or rapid swinging when moving heavy objects in a confined space, which are dynamic processes with high-risk characteristics.

[0079] The system utilizes a built-in template library of electrical violation behaviors, which pre-stores standard instantaneous acceleration rate of change sequences and standard voxel data sequences for various typical violation scenarios. These standard sequences are derived from baseline features obtained through data cleaning and modeling of a large number of historical violation cases (such as panicked actions caused by operating without disconnecting power, jumping movements in violation of safety regulations, and tool throwing) and normal operational data. The system calculates the Euclidean distance between the instantaneous acceleration rate of change sequence and each standard violation sequence in the template library in real time to quantify the geometric difference in the dynamic trajectory between the actual operation and the standard violation pattern. To eliminate the influence of different operation intensities and sensor ranges, the system normalizes the calculated Euclidean distance, mapping it to the interval between 0 and 1, and finally generates a dynamic matching score.

[0080] This score intuitively reflects the similarity between the target personnel's actual movement pattern and a specific violation: the lower the score (the smaller the normalized distance), the closer the mechanical characteristics of the actual movement are to the preset serious violation template (such as violent operation or uncontrolled fall), and the higher the probability of being judged as a violation; conversely, the higher the score, the smoother the movement and the more compliant it is with safe operating procedures. Through this process, abstract work postures can be transformed into quantifiable violation risk indicators from the microscopic dynamic characteristics level, realizing automated and high-precision identification of non-driving violations in complex power scenarios such as substation switching operations and line maintenance, effectively making up for the shortcomings of traditional video surveillance, which can only identify static postures and is difficult to capture instantaneous dynamic risks.

[0081] Step 63: If the dynamic matching score is greater than the third threshold, then mark the spatiotemporal behavior voxel data as a first-level confidence state; otherwise, proceed to step 64.

[0082] Step 64: Obtain the planar coordinate sequence of the virtual grid node where the target character is located in five consecutive frames, and construct a discrete path point set; select three consecutive points in the path point set, calculate the absolute value of the difference between the slope of the line connecting the first two points and the slope of the line connecting the last two points, and use it as the local curvature value; obtain the normalized path curvature index based on the local curvature value; if the normalized path curvature index is greater than the fourth threshold, mark the spatiotemporal behavior voxel data as a second-level confidence state; otherwise, remove the spatiotemporal behavior voxel data.

[0083] The third threshold is determined based on the statistical distribution of the stability of normal operating actions in power operations. This threshold needs to cover the dynamic fluctuation range under more than 95% of the standard operating scenarios. In this embodiment, it is set to 0.72. If the calculated dynamic matching score is greater than 0.72, it indicates that the target person's motion characteristics are significantly different from the standard violation template, belonging to normal or low-risk operating behavior. The system will directly mark the corresponding spatiotemporal behavior voxel data as a first-level confidence state and enter the subsequent high-precision analysis process. Conversely, if the score is less than or equal to 0.72, step 64 is triggered to perform in-depth path geometric feature verification.

[0084] In step 64, the planar coordinate sequence of the virtual grid node where the target person is located within a continuous five-frame time window is extracted to construct a discrete path point set to represent its movement trajectory. Any three consecutive adjacent points in the path point set are selected, and the slopes of the lines connecting the first two points and the last two points are calculated. The absolute value of the difference between these slopes is taken as the local curvature value. This index is used to quantify the degree of curvature and the intensity of directional changes at that point. Subsequently, combining the geometric characteristics of common walking, climbing, and turning movements in power operation sites, the local curvature value is mapped to the interval between 0 and 1 to obtain a normalized path curvature index. For the fourth threshold, it is set based on the sharp path characteristics presented by typical violations (such as sharp turns when illegally crossing warning lines, abnormal backtracking, or blind detours under equipment obstruction). This threshold needs to be able to distinguish between normal meandering movement and abnormal sharp turns; in this embodiment, it is set to 0.85.

[0085] If the normalized path curvature index is greater than 0.85, it indicates that the target person underwent a drastic change in direction or trajectory distortion within a very short period of time, which conforms to the kinematic characteristics of certain specific violation scenarios (such as suddenly entering a live area or making a rapid illegal turnback). However, considering the possibility of sensor jitter or transient interference, the data is not directly discarded. Instead, the spatiotemporal behavior voxel data is marked as a level 2 confidence state, indicating that the data has abnormal characteristics but has further analytical value and needs to be confirmed again in conjunction with the context. If the normalized path curvature index is less than or equal to 0.85, and the previous dynamic matching score has been lower than the third threshold, then it is determined that the data segment does not conform to the dynamic characteristics of normal operation, and its path shape also lacks typical characteristics of sudden violation. It is very likely false data caused by environmental noise, signal loss, or invalid background clutter. The data segment will be discarded and will no longer be included in the subsequent violation judgment model, thereby effectively purifying the input data and improving the overall system's anti-interference ability and judgment accuracy.

[0086] Step 65: Integrate the first voxel data, the spatiotemporal behavioral voxel data, and the corresponding confidence labels into a voxel data sequence.

[0087] First, the system retrieves the first voxel data, the original spatiotemporal behavioral voxel data of the target character acquired and associated, and the confidence labels (divided into Level 1 confidence, Level 2 confidence, or removal labels) determined by the dynamic matching score and path curvature index. These three types of data are strictly aligned according to timestamps and encapsulated into a structured voxel data sequence frame by frame. In terms of data structure design, each sequence unit not only contains the voxel mesh matrix in 3D space and its corresponding texture gradient variance value, but also embeds dynamically updated confidence metadata: if a frame is labeled as Level 1 confidence, its metadata includes a high dynamic matching score and a low path curvature index, indicating that the frame data is highly reliable and conforms to normal operating standards; if labeled as Level 2 confidence, it records a lower matching score and a higher local curvature value, indicating that the frame has abnormal motion characteristics but still retains analytical value after correction; for data deemed invalid, the system directly removes it from the sequence or marks it as null to avoid interfering with subsequent model training. The final generated voxel data sequence forms a timeline with quality grading labels.

[0088] Step 66: Generate the source tracing result based on the confidence labels of the voxel data sequence.

[0089] In one specific embodiment of this application, effective segments (i.e., data labeled with first- or second-level confidence levels) are first extracted from the voxel data sequence. The mean Sobel gradient magnitude and average position vector of the sequence to be tested are directly input into the corresponding standard sequences pre-stored in the power violation behavior template library for point-by-point comparison. The matching status is determined by calculating the numerical similarity between the two in the two dimensions of mean Sobel gradient magnitude and average position vector. If the similarity of both parameters falls within the preset high similarity interval (e.g., gradient similarity > 0.85 and overall position vector similarity > 0.90), then the match is considered successful. Once a match is successful, a differentiated source tracing result generation mechanism is immediately activated: If the original data is marked with a Level 1 confidence level, the source tracing result will clearly indicate that the violation has extremely high physical authenticity, and will include a voxel deviation heatmap of the standard sequence and the actual sequence at key violation action nodes (such as hand position, tool angle), visually demonstrating the high degree of overlap between the measured data and the standard model; if marked with a Level 2 confidence level, the source tracing result will focus on pointing out the differences between the abnormal trajectory and the standard model (such as non-standard voxel distribution caused by abrupt changes in path curvature), suggesting that the violation may be accompanied by sudden evasive actions or environmental interference, while also marking the specific parameter fluctuation frames that caused the confidence level to decrease. The final source tracing result not only confirms the violation type and specific similarity score, but also provides a quantifiable chain of evidence through an intuitive standard-measured voxel overlay comparison map.

[0090] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for tracing the source of power-related violations by integrating panoramic images and location data, characterized in that, Includes the following steps: Step 1: Obtain the panoramic monitoring image sequence and the original location data stream of the positioning system; the panoramic monitoring image sequence includes an initial frame image arranged in chronological order and several subsequent frames. Step 2: Construct an initial two-dimensional topological mesh model based on the initial frame image. The initial two-dimensional topological mesh model is composed of several triangular mesh units. Each triangular mesh unit is defined by three virtual mesh nodes, and each virtual mesh node includes planar coordinates. Step 3: Based on several frames of subsequent images, select the current frame image to be processed in sequence, parse the original position data stream, extract the three-dimensional physical coordinates of the positioning device under the preset timestamp, and map the three-dimensional physical coordinates to the initial two-dimensional topological mesh model through projection transformation to obtain the two-dimensional projection position of the spatial anchor point; Step 4: Based on the spatial anchor point, the two-dimensional projection position is mapped to the initial two-dimensional topological mesh model, and the triangular mesh cell containing the two-dimensional projection position is identified as the target cell; the three virtual mesh nodes of the target cell are respectively used as the first local control node, the second local control node, and the third local control node; the arithmetic mean of the planar coordinates of the first local control node, the second local control node, and the third local control node is used as the geometric center, and the deviation vector is determined according to the geometric center of the target cell and the two-dimensional projection position; Step 5: Obtain the real-time area distortion rate of the target unit and the topology tearing risk factors of the first local control node, the second local control node and the third local control node, and generate correction coefficients together. Generate an effective displacement correction amount based on the correction coefficients and the deviation vector, and generate a corrected topology mesh model for each frame based on the effective displacement correction amount. Step 6: Based on the corrected topological grid model, acquire the spatiotemporal behavioral voxel data of all frames, execute a multi-level verification process to generate a voxel data sequence with confidence labels, and perform multi-dimensional feature matching in a preset violation behavior template library to output the violation source tracing result.

2. The method for tracing power violations by fusing panoramic images and positioning data according to claim 1, characterized in that, The real-time area distortion rate of the target cell and the topology tearing risk factors of the first local control node, the second local control node, and the third local control node are obtained, and correction coefficients are generated together. An effective displacement correction is generated based on the correction coefficients and the deviation vector. A corrected topology mesh model for each frame is generated based on the effective displacement correction. This process includes the following steps: Step 51: Obtain the historical reference area and the instantaneous area of ​​the target unit in the current frame, and obtain the real-time area distortion rate based on the instantaneous area and the historical reference area; Step 52: Obtain the first plane coordinate set of the first local control node, the second local control node and the third local control node in the previous frame, and the second plane coordinate set in the current frame; Step 53: Based on the first plane coordinate set and the second plane coordinate set, determine the position vectors of the first local control node, the second local control node, and the third local control node; based on the first plane coordinate set and the second plane coordinate set of the first local control node, the second local control node, and the third local control node, determine the relative scaling ratio of each side length of the target unit; Step 54: Calculate the cosine of the angle between adjacent position vectors; determine the topology tear risk factor based on the cosine of the angle and the relative scaling ratio; Step 55: If the real-time area distortion rate is greater than the preset area distortion safety threshold, then a correction coefficient is constructed based on the preset area distortion safety threshold and the real-time area distortion rate; if the topology tearing risk factor is greater than the preset tearing risk safety threshold, then the correction coefficient is defined as 0; otherwise, the preset base coefficient is used as the correction coefficient. Step 56: Generate an effective displacement correction amount based on the correction coefficient and the deviation vector, and correct the first local control node, the second local control node, the third local control node and the virtual mesh node based on the effective displacement correction amount to obtain a corrected topological mesh model.

3. The method for tracing power violations by fusing panoramic images and positioning data according to claim 2, characterized in that, Based on the planar coordinates of the first local control node, the second local control node, and the third local control node, and the effective displacement correction amount, the first optimized node coordinates are obtained; based on the virtual mesh nodes inside the target unit, the first effective displacement correction amount is constructed through the effective displacement correction amounts of the first local control node, the second local control node, and the third local control node, and the first effective displacement correction amount is superimposed on the planar coordinates based on the virtual mesh nodes inside the target unit to obtain the second optimized node coordinates.

4. The method for tracing the source of power violations by fusing panoramic images and positioning data according to claim 3, characterized in that, Based on the virtual mesh nodes of the triangular mesh units other than the target unit, search for the N nearest corrected nodes, where the corrected nodes are the first local control node, the second local control node, the third local control node, and the virtual mesh node of the target unit; Determine the first distance between the virtual mesh nodes of the triangular mesh elements outside the target element and the corrected nodes. Determine the normalized weight coefficient based on the first distance. Obtain the coordinates of the third optimized node based on the effective displacement correction amount of the corrected nodes and the normalized weight coefficient.

5. The method for tracing the source of power violations by fusing panoramic images and positioning data according to claim 4, characterized in that, Based on the coordinates of the first optimized node, the second optimized node, and the third optimized node, the corrected topology mesh model is obtained.

6. The method for tracing power violations by fusing panoramic images and positioning data according to claim 2, characterized in that, A multi-level verification process is executed to generate a voxel data sequence with confidence labels, and multi-dimensional feature matching is performed in a pre-defined violation behavior template library to output the violation tracing results, including the following steps: Step 61: Based on the spatiotemporal behavior voxel data, obtain the texture gradient variance value and the signal-to-noise ratio of the positioning signal of the triangular mesh cell where the target character is located; if the texture gradient variance value is less than the first threshold or the signal-to-noise ratio of the positioning signal is less than the second threshold, it is determined that there is environmental interference in the current spatiotemporal behavior voxel data, and the spatiotemporal behavior voxel data is processed by compensation branch: the spatiotemporal behavior voxel data of adjacent frames are weighted and smoothed to generate the first voxel data and mark it as the second confidence state; otherwise, the spatiotemporal behavior voxel data is transmitted to step 62 to execute the first level judgment process; Step 62: Determine the instantaneous acceleration vector based on the position vector, and obtain the instantaneous acceleration rate of change based on the instantaneous acceleration vector; obtain the standard instantaneous acceleration rate of change sequence based on the preset violation behavior template library, calculate the Euclidean distance between the instantaneous acceleration rate of change and each standard instantaneous acceleration rate of change in the standard instantaneous acceleration rate of change sequence, and normalize the Euclidean distance to obtain the dynamic matching score; Step 63: If the dynamic matching score is greater than the third threshold, then mark the spatiotemporal behavior voxel data as a first-level confidence state; otherwise, proceed to step 64. Step 64: Obtain the planar coordinate sequence of the virtual grid node where the target character is located in five consecutive frames, and construct a discrete path point set; select three consecutive points in the path point set, calculate the absolute value of the difference between the slope of the line connecting the first two points and the slope of the line connecting the last two points, and use it as the local curvature value; obtain the normalized path curvature index based on the local curvature value; if the normalized path curvature index is greater than the fourth threshold, then mark the spatiotemporal behavior voxel data as a second-level confidence state; otherwise, then remove the spatiotemporal behavior voxel data. Step 65: Integrate the first voxel data, the spatiotemporal behavioral voxel data, and the corresponding confidence labels into a voxel data sequence; Step 66: Generate the source tracing result based on the confidence labels of the voxel data sequence.

Images