Near-electricity operation risk dynamic identification method fusing fine segmentation and binocular vision

Abstract

The invention provides a near-electricity operation risk dynamic identification method fusing fine segmentation and binocular vision. The near-electricity operation risk dynamic identification method comprises the steps that a left view image, shot by a binocular camera, of substation near-electricity operation is acquired; initializing calibration parameters of the segmentation model, the binocular three-dimensional distance measurement model and the binocular camera; identifying and extracting key job elements in the left view image through a segmentation model, wherein a segmentation result comprises each detected instance and a two-dimensional position thereof; for each instance, the three-dimensional position of the instance in the space is predicted through a binocular three-dimensional ranging model, and the three-dimensional coordinates of the instance are obtained; according to the three-dimensional coordinates of each instance, the Euclidean distance between different key operation elements is calculated, and the minimum distance between the key operation elements is obtained; according to the invention, whether there is an operation risk is determined according to the minimum distance between different key operation elements, and the method has high distance measurement precision and real-time performance, and is helpful for improving the safety of near-electricity operation and the risk early warning capability.

Description

Technical Field

[0001] The present invention relates to the technical field of operation risk identification, and in particular to a method for dynamic identification of near-electrical operation risks integrating fine segmentation and binocular vision. Background Art

[0002] The safe distance near the live wire is one of the key reasons for electric shock accidents. Its concept comes from GB26860-2011 "Electric Power Safety Work Regulations (Electrical Part of Power Plants and Substations)" issued by the State Grid. According to the safe distance near the live wire corresponding to different voltage levels and the restrictions of work tickets, the minimum air interval between the operator and the live wire should be maintained, and the interval margin should be reserved to ensure sufficient safety. The near-live wire operation site of the substation is usually complex and diverse, dynamically coherent, highly instantaneous, and highly correlated between different scenes. The current operation distance risk identification mainly relies on auxiliary control means such as positioning systems, physical sensors and video monitoring, providing key information for manual decision-making and processing, and providing a solid digital identification technology foundation for the risk control of live wire operations in substations, but there are still many limitations. Positioning systems and physical sensors are usually limited by accuracy, and additional complex operation requirements are imposed on front-line operators, and the visualization of key information is poor. Video surveillance is more common in the distance control of near-electrical operations, and is more accepted by operators and supervisors due to its visual intuitiveness, with a high coverage rate. However, this type of simple visual technology lacks the ability to self-process rich scene information, which makes manual "one-to-many" verification more complex and repetitive. This may cause operators to make subjective misjudgments of distance information, resulting in efficiency and fatigue problems. Summary of the invention

[0003] The purpose of the present invention is to provide a method for dynamic identification of risks of near-electrical operations that integrates fine segmentation and binocular vision. The method achieves the identification of safety risks of near-electrical operations in substations through the coordination of segmentation positioning and stereo perception.

[0004] To achieve the above-mentioned purpose, the present invention provides a method for dynamic identification of risks of near-electrical operations by integrating fine segmentation and binocular vision, the method comprising the steps of:

[0005] S101, obtaining a left view image of a substation near-power operation taken by a binocular camera;

[0006] S102, initializing the segmentation model, the binocular 3D ranging model and the calibration parameters of the binocular camera;

[0007] S103, identifying and extracting key operating elements in the left view image through a segmentation model, wherein the segmentation result includes each detected instance and its two-dimensional position;

[0008] S104, for each instance, predicting its three-dimensional position in space by using a binocular three-dimensional ranging model to obtain the three-dimensional coordinates of the instance;

[0009] S105, calculating the Euclidean distances between different key operation elements according to the three-dimensional coordinates of each instance, and obtaining the minimum distance between the key operation elements;

[0010] S106. Determine whether there is an operation risk based on the minimum distance between different key operation elements.

[0011] Furthermore, the segmentation model adopts the YOLOv8 segmentation model.

[0012] Furthermore, the DCNv4-Dyhead detection head is used to replace the baseline model detection head of the YOLOv8 segmentation model. The DCNv4-Dyhead detection head replaces the 3×3 convolution in the Dyhead spatial attention with DCNv4 (DeformableConvNets v4), and removes unnecessary softmax normalization in spatial aggregation. The spatial attention π of the DCNv4-Dyhead detection head S (·)The calculation formula is as follows:

[0013]

[0014] In the above formula, the feature tensor F∈R L×S×C , S represents the spatial location information, C represents the channel information, L is the number of input feature layer levels, G is the total number of clustering groups, K is the number of sparse sampling locations, and w l,g represents the weight parameter, which is used to control the weight relationship between the lth channel and the gth group, F(l; p k +Δp gk ; c) indicates that in the lth channel, position p k After an offset of Δp gk The offset is used to calculate the dynamic convolution kernel, c represents the channel dimension, m represents the feature map value after gk represents the modulation scalar of the k-th sampling point in the g-th group, normalized along the K dimension using the softmax function.

[0015] Further, introduce the Layer-Adaptive Magnitude-based Pruning (LAMP) method into the segmentation model. Determine the pruning priority by calculating the LAMP scores of each layer's weights to maintain the balance between model performance and sparsity. Specifically, it includes: flatten each weight tensor into a one-dimensional vector, sort it in ascending order according to the given index map. Whenever u < v holds and |W[u]| ≤ |W[v]|, where W[u] represents the u-th element in the weight tensor, the LAMP score for this element is defined as:

[0016]

[0017] u represents the index of the weight element currently being calculated, and v represents the index of other weight elements relative to u.

[0018] Further, for each instance, predict its three-dimensional position in space through the binocular three-dimensional ranging model to obtain the three-dimensional coordinates of the instance. Specifically, it includes the following steps:

[0019] S201. Calculate a high-quality disparity map through AANet (Adaptive Aggregation Network) stereo matching, and convert the disparity map into a depth map that associates pixel coordinates (u, v) with Z-axis information. The conversion relationship is:

[0020]

[0021] In the above formula, Z is the depth value, f is the camera focal length, b is the baseline distance of the binocular camera, and disp is the disparity value;

[0022] S202. Combine the depth map and the internal parameters of the binocular module to reconstruct the real point coordinates (X, Y, Z) of the corresponding pixel points in the world coordinate system, so as to correspond the two-dimensional pixel positions in the image to the three-dimensional positions in the actual scene. The calculation formula is:

[0023]

[0024] where c x ,c y represents the principal point position;

[0025] S203. Within the effective depth range of the same key operation element, select multiple points with depth values and spans that meet the requirements for depth assignment;

[0026] S204. Calculate the Euclidean distance between the nearest depth assignment point pairs of two key operation elements, and return the minimum distance between the key operation elements.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. Aiming at the complex environment of substation scenes, this paper innovatively replaces the original YOLOv8 benchmark model detection head with the DCNv4-Dyhead detection head, significantly enhancing the model's ability to learn features such as the proportion, position, and deformation of flexibly changing elements. This replacement not only improves the fine-grained segmentation performance of the model, but also demonstrates superiority in multi-scale feature learning;

[0029] 2. The present invention adopts the LAMP pruning method to effectively reduce the number of model parameters and computational complexity, ensuring that the model is lightweight and improves reasoning speed while maintaining high accuracy.

[0030] 3. Through the effective integration of binocular vision technology, the present invention can accurately analyze the safe distance between key operating elements in complex operating scenes, thereby significantly improving the safety and risk warning capabilities of near-electrical operations. The method has high ranging accuracy and real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0032] Figure 1 It is a schematic diagram of the overall process of a method for dynamic identification of risks of near-electrical operations integrating fine segmentation and binocular vision provided by an embodiment of the present invention.

[0033] Figure 2 It is a schematic diagram of the structure of the DCNv4-Dyhead detection head provided in an embodiment of the present invention.

[0034] Figure 3 It is a schematic diagram of the segmentation model pruning process provided by an embodiment of the present invention.

[0035] Figure 4 3 is a schematic diagram for comparing stereo matching models provided by the experimental example of the present invention.

[0036] Figure 5 It is a comparison chart of the number of model pruning channels provided by the experimental example of the present invention.

[0037] Figure 6 It is a comparison diagram of the model PR curve provided by the experimental example of the present invention.

[0038] Figure 7 It is a comparison diagram of the model plane fine division effect provided by the experimental example of the present invention.

[0039] Figure 8 It is a visual comparison diagram of the dynamic identification model of substation near-electricity operation risk provided by the experimental example of the present invention. DETAILED DESCRIPTION

[0040] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0041] Example

[0042] Reference Figure 1 This embodiment provides a method for dynamic identification of risks of near-electrical operations by integrating fine segmentation and binocular vision, the method comprising the following steps:

[0043] S101, obtaining a left view image of a substation near-power operation taken by a binocular camera.

[0044] S102, initializing the segmentation model, the binocular 3D ranging model and the calibration parameters of the binocular camera.

[0045] S103, identifying and extracting key operating elements in the left view image through the segmentation model, and the segmentation result includes each detected instance and its two-dimensional position.

[0046] S104: For each instance, predict its three-dimensional position in space through a binocular three-dimensional ranging model to obtain the three-dimensional coordinates of the instance.

[0047] S105. Calculate the Euclidean distances between different key operation elements according to the three-dimensional coordinates of each instance, and obtain the minimum distance between the key operation elements.

[0048] S106. Determine whether there is an operation risk based on the minimum distance between different key operation elements.

[0049] As a preferred implementation, the segmentation model uses the YOLOv8 segmentation model. The YOLO series of models include image classification, target detection, instance segmentation, and posture key point detection, and a unified neural network can predict multiple outputs, such as bounding boxes, category probabilities, and object masks.

[0050] In view of the complexity and diversity of near-electric operation scenarios, the YOLOv8 segmentation model used in this embodiment generates a pixel-level mask for the operation target for 2D division and positioning. Compared with the target detection task, it can avoid the interference of non-target elements in the detection box and provide more accurate spacing calculation for multi-point and large-span depth assignment processing in the same instance.

[0051] Since key operating elements such as personnel and vehicles in near-electric operation scenes appear in the form of coherent features such as dynamics, interactions, and dependencies, in order to better accurately divide and locate such scenes, this embodiment replaces the baseline model detection head of the original YOLOv8 segmentation model with the DCNv4-Dyhead detection head to capture the inherent spatial relationship between different scales and shapes of different key operating elements, and enhance the learning ability of multi-scale features such as the proportion, position, and deformation of flexibly changing elements.

[0052] Dynamic Head (Dyhead) is a dynamic head structure, such as Figure 2 As shown in the figure, it can adaptively adjust the weights according to the input features, so that each convolution kernel can dynamically adjust the weights according to the input features, thereby improving the adaptability to complex scenes and targets. Specifically, the feature dimensions L, S, and C represent different scales (and feature maps of different layers or stages), spatial position information (S = H × W), and channel information, respectively. The calculation formula of attention W(·) is as follows:

[0053] W(F)=π C (π S (π L (F)·F)·F)·F

[0054] In the above formula, the feature tensor F∈R is given on the detection layer L×S×C , through three different attention mechanisms - scale attention π L (·), spatial attention π S (·) and channel attention π C (·) to achieve dynamic adjustment and aggregation of features.

[0055] Scaled Attention π L (·) Fusion of features of different scales based on their semantic importance is achieved through a combination of global pooling, convolution, and activation functions. Figure 2 As shown in the figure, Adaptive Avg Pool 2d is used to perform global pooling on the height H and width W dimensions of the feature map to obtain the global features of each channel. All channels are integrated through a 1×1 convolution to output L feature channels. Finally, ReLU and hard sigmoid activation functions are applied to enhance the nonlinear representation ability of the features. The calculation formula of scale attention is as follows:

[0056]

[0057] In the above formula, f(·) is a linear function, which is approximated by 1×1 convolution, and σ(x) is a hard sigmoid activation function. The scale attention module is deployed on the scale dimension of the feature to learn the importance of each scale feature, thereby enhancing the features at the appropriate scale to adapt to the size of the individual target.

[0058] Spatial attention focuses on the discrimination ability of different spatial positions. Considering the high dimensionality of S, it is decoupled, and variable convolution is used to sparse the attention learning, and then the features are integrated across scales.

[0059] The sampling points of traditional convolution are fixed, but deformable convolution dynamically adjusts the position of the sampling points through the offset in the convolution operation to achieve more accurate capture of irregular objects. The offset is learned through an additional convolution layer and then used to correct the position of each sampling point during the convolution process to complete the deformable convolution.

[0060] This embodiment replaces the 3×3 convolution in the original Dyhead spatial attention with Deformable ConvNetsv4 (DCNv4). This efficient deformable convolution operator optimizes memory access based on the predecessor DCNv3, reducing unnecessary computational overhead by nearly half. At the same time, it removes unnecessary softmax normalization in spatial aggregation to enhance its dynamic characteristics and expression ability. Improved DCNv4-Dyhead spatial attention π S The calculation formula of (·) is as follows:

[0061]

[0062] In the above formula, the feature tensor F∈R L×S×C , S represents the spatial location information, C represents the channel information, L is the number of input feature layer levels, G is the total number of clustering groups, K is the number of sparse sampling locations, and w l,g represents the weight parameter, which is used to control the weight relationship between the lth channel and the gth group, F(l; p k +Δp gk ; c) indicates that in the lth channel, position p k After an offset of Δp gk The offset is used to calculate the dynamic convolution kernel, c represents the channel dimension, m represents the feature map value after gk represents the modulation scalar of the k-th sampling point in the g-th group, normalized along the K dimension using the softmax function.

[0063] Channel Attention π C(·) Dynamically switch feature channels to assist different tasks, promoting joint learning and generalization of target expression capabilities. Channel attention models and adjusts channels through two fully connected layers, so that information between different channels can be more effectively integrated and expressed. This improves the task perception ability of the model. Channel attention π C The calculation formula of (·) is as follows:

[0064] π C (F)·F=max(α 1 (F)·F c +β 1 (F),α 2 (F)·F c +β 2 (F)

[0065] In the above formula, F c is the characteristic slice of the cth channel, [α 1 ,α 2 ,β 1 ,β 2 ] T =θ(·) is a hyperparameter used to control the activation threshold, similar to DyReLU.

[0066] DCNv4-DyHead stacks three attention mechanisms in one block, each of which contains scale perception, spatial and channel attention. The final segmentation model detection head is composed of multiple such blocks, providing better feature representation and recognition capabilities for key operational elements in risk management tasks.

[0067] As another preferred implementation, this embodiment starts from the power visual edge intelligence and considers completing the model recognition processing at the edge side where the data is generated. On the basis of the previous embodiment, this embodiment further prunes the model to ensure the response speed of the model to cope with the strong instantaneous nature of the risk of near-electrical operation scenarios.

[0068] Pruning is a method that aims to improve the efficiency and inference speed of neural network models by reducing the number of model parameters. Traditional pruning methods usually adopt a uniform strategy, which tends to ignore the importance differences of parameters between different layers, but in exchange for computing space, it leads to a significant decrease in model accuracy. This implementation introduces a layer-adaptive magnitude-based pruning method (LAMP) in the segmentation model, and determines the pruning priority by calculating the LAMP score of each layer weight, thereby maintaining a balance between model performance and sparsity. Figure 3As shown, the amplitude pruning method specifically includes: flattening each weight tensor into a one-dimensional vector, sorting in ascending order according to a given index map, and u<v holds, |W[u]|≤|W[v]|, where W[u] represents the u-th element in the weight tensor, and the LAMP score for this element is defined as:

[0069]

[0070] u represents the index of the weight element currently being calculated, and v represents the index of other weight elements relative to u. The LAMP score is a rescaled version of the absolute value of the amplitude. The global pruning is performed according to the LAMP score, and the weights with high LAMP scores are retained until the required global sparsity constraint is met. There is a connection with a LAMP score of 1 in each layer to ensure that each layer retains at least one surviving connection to avoid layer collapse.

[0071] The aforementioned segmentation model accurately locates and finely divides the two-dimensional images of key operating elements in the substation near-electric operation scene, such as personnel, vehicles, transformers, etc. in complex environments. On this basis, this embodiment further constructs a near-electric spatial distance perception model based on binocular vision to accurately measure the spatial distance of key operating elements for risk identification.

[0072] Binocular vision technology simulates the stereoscopic perception of human eyes and uses two cameras to capture images of the same scene from different angles. The geometric accuracy of the imaging is ensured by calibrating the internal parameters of the camera (including focal length, imaging origin, distortion coefficient, etc.), and the relative position of the binocular camera system is calibrated at the same time, and the spatial relationship between the two cameras is described by the rotation matrix and displacement vector. Exemplarily, the internal and external parameters of the binocular camera system can be obtained by Zhang Zhengyou calibration method or other methods, which are not specifically limited in this embodiment.

[0073] Stereo matching is the key to binocular vision accuracy. The stereo matching algorithm is used to establish the feature relationship between the left and right views. By matching the corresponding image points, the disparity map is converted into a depth map with spatial information to calculate the distance. In order to improve the accuracy and real-time performance of stereo matching in the substation near-electric operation scene, this embodiment introduces AANet (Adaptive Aggregation Network), a stereo matching method based on deep learning, into the near-electric spatial distance perception model. This method achieves efficient disparity estimation in complex environments through multi-scale feature extraction and adaptive aggregation strategies. A feature extractor with shared weights is used to extract features from the left and right images, and three feature maps with different resolutions (1 / 3, 1 / 6, and 1 / 12) are constructed through a spatial pyramid structure, thereby enhancing the feature expression capability at multiple scales. Different from the traditional 4D cost volume construction method, AANet uses the correlation calculation between the left and right image features to construct a 3D cost volume, which significantly reduces the computational complexity, as shown in the following formula:

[0074]

[0075] In the formula, C s (d,h,w) represents the matching cost of the disparity candidate value d at the pixel position (h,w), calculated The inner product of the feature vectors of the two left and right images. The larger the inner product value, the more similar the two features are. The matching cost C s The smaller (d,h,w) is, the similar calculation is performed on each disparity value to construct a three-dimensional cost volume for further depth estimation.

[0076] In the cost aggregation stage, the adaptive intra-scale aggregation module (ISA) uses the feature representation of sparse points and deformable convolution to achieve more flexible cost aggregation, overcoming the limitations of traditional local cost aggregation in processing discontinuous areas of disparity. Specifically, the aggregation formula in the ISA module is as follows:

[0077]

[0078] in, represents the aggregation cost of pixel p at disparity candidate d, K 2 is the number of sampling points, w k is the aggregation weight of the kth sampling point, p k is a fixed offset relative to the pixel, and the offset Δp is extracted from the input cost volume through a separate convolutional layer k and weight m k ,Achieve adaptive sampling and content-adaptive cost aggregation.

[0079] The adaptive cross-scale aggregation module (CSA) is used to aggregate cost volume features of different scales. It combines coarse scale and detail features, and enhances the disparity prediction ability of textureless areas and subtle structures through downsampling and upsampling operations. The aggregation formula of the CSA module is as follows:

[0080]

[0081] In the above formula, is the result cost body after cross-scale cost aggregation, is the same-scale aggregate cost volume at the k-th scale, f k is a function used to adaptively combine cost volumes of different scales. Lower resolution scales can capture more global structural information, while high resolution scales are better at capturing detailed features. Therefore, through cross-scale aggregation, the CSA module can utilize coarser scale information in textureless or weakly textured areas while retaining higher resolution in detail-rich areas, thereby balancing the relationship between global information and detail information.

[0082] Specifically, for each instance, the three-dimensional position of the instance in space is predicted by the binocular three-dimensional ranging model to obtain the three-dimensional coordinates of the instance, including the following steps:

[0083] S201, calculate a high-quality disparity map through AANet (Adaptive Aggregation Network) stereo matching, and convert the disparity map into a depth map that stores Z-axis information in association with pixel coordinates (u, v). The conversion relationship is:

[0084]

[0085] In the above formula, Z is the depth value, f is the focal length of the camera, b is the baseline distance of the binocular camera, and disp is the disparity value.

[0086] like Figure 4 As shown, the disparity map and depth map generated by the method provided in this embodiment are better in detail depiction and smoothness of depth changes than the disparity map and depth map generated by the classic stereo matching network PSMNet. At the edges of objects and in complex texture areas, the method can better preserve edge details and texture information, avoiding parallax blur and depth discontinuity.

[0087] S202, combining the depth map and the internal parameters of the binocular module, reconstructing the real point coordinates (X, Y, Z) of the corresponding pixel point in the world coordinate system, so as to correspond the two-dimensional pixel position in the image to the three-dimensional position in the actual scene, and the calculation formula is:

[0088]

[0089] where c x , c y Indicates the principal point position.

[0090] S203. Within the effective depth range of the same key operating element, select multiple points whose depth values ​​and spans meet the requirements for depth assignment.

[0091] After obtaining the three-dimensional position coordinates, the model provided in this embodiment selects multiple points with very stable depth values ​​and large spans within the effective depth range of the same key operating element for depth assignment, rather than using all points, to reduce computing consumption. For example, the number of depth assignment points in the segmentation task can be 3.

[0092] S204, calculate the Euclidean distance of the closest depth assignment point pair between two key operating elements, and return the minimum distance between the key operating elements. Compared with the single-point random or geometric center depth assignment method, it is more accurate and effectively reduces errors when processing complex structures or irregularly shaped target objects. The distance D calculation formula is as follows:

[0093]

[0094] In the above formula, (x i ,y i ,z i ) and (z j ,y j ,z j ) represent the three-dimensional position coordinates of two points respectively.

[0095] Experimental example

[0096] Since there is currently no typical open source dataset for substation near-electrical operation scenes. For this reason, this experimental example constructs a special substation near-electrical operation scene dataset. This dataset contains 6001 visible light images of key elements of near-electrical operations such as workers, vehicles, and transformers. These images come from substations of different voltage levels. Through multiple acquisitions and careful annotation, they cover a variety of scenes with multiple angles and multiple element proportions. Table 1 shows that the 6001 images are divided into training sets, validation sets, and test sets in a ratio of 6:2:2. The left view in the binocular visible light image is used for the training and experiment of the plane operation element division model, while the monocular visible light image does not participate in the training and testing of the stereo matching network in the binocular ranging model.

[0097] Table 1 Composition of substation near-power operation scenario dataset

[0098]

[0099] In this experiment, the segmentation model and the stereo matching network in the binocular 3D ranging model are trained on a high-performance computing platform. The detailed parameters of the experimental training platform are shown in Table 2.

[0100] Table 2 Experimental platform parameters

[0101]

[0102] In this experimental example, the training image input size of the segmentation model is 640×640, and the batch size is 16, in order to effectively balance the training and computational load. The model uses the SGD optimizer, with a momentum of 0.9, L2 regularization to avoid overfitting, and a weight decay coefficient of 0.0005. During the training process, the learning rate adopts a piecewise decay strategy, with an initial learning rate of 0.01 and a decay factor of 0.1. The training is performed for a total of 150 epochs.

[0103] In order to comprehensively evaluate the performance of the segmentation model, five key indicators were selected: mAP50, mAP50-95, model size, GFLOPs, and FPS.

[0104] mAP50 and mAP50-95 are the core indicators for measuring the accuracy of target detection. mAP50 refers to the average accuracy under 50% IoU conditions, and mAP50-95 refers to the average accuracy between 50% and 95% IoU. The IoU of the instance segmentation task is the degree of overlap between the predicted mask and the true mask. When the IoU is greater than the set threshold, the prediction result is positive. The calculation formulas for precision P and recall R are as follows:

[0105]

[0106] In the formula, TP is the number of correctly predicted targets, FP is the number of other categories that are mistakenly identified as targets, and FN is the number of targets that are mistakenly identified as other categories.

[0107] The area enclosed by the PR (Precision-Recall) curve of each category is AP (average precision). mAP (mean average precision) is the average of the AP of all categories. The higher the value, the more balanced and superior the model's performance in multi-category plane fine segmentation tasks, as shown in the following formula:

[0108]

[0109] Model Size is the size of the weight file generated by training. A smaller model size means that the model's network structure is more refined and optimized, reducing unnecessary parameter redundancy while maintaining high performance. GFLOPs (Giga Floating Point Operations per Second) is a standard for measuring the computational complexity of a model, indicating the number of floating-point operations performed per second. A lower GFLOPs value indicates that the model requires less computing resources on edge computing devices. FPS (Frames Per Second) indicates the number of frames that a model can process per second. These three parameters can be used to compare the model's network refinement, computational complexity, and real-time processing capabilities.

[0110] This experiment compares the effect of connecting different numbers of DCNv4-Dyhead modules in series on the performance of the YOLOv8s segmentation model. By comparing the series connection experiments of 1 to 5 modules, the performance indicators such as mAP50, model weight file size, GFLOPs, and FPS are listed in Table 3.

[0111] For each additional DCNv4-Dyhead module, the GFLOPs of the model increased by about 0.3, indicating that the module has a small computational burden. Compared with DCNv3-Dyhead, DCNv4 significantly reduces the consumption of redundant computing resources, so there is no significant loss in inference speed when adding multiple modules. When two DCNv4-Dyhead modules are connected in series, the mAP50 of the model increases from the baseline 0.825 to 0.836, an increase of 0.011. This shows that the accuracy of the model has been improved without a significant increase in the amount of calculation. Under the configuration of two modules, the model's accuracy, computational burden, and inference speed perform best. After increasing to three modules, the increase in mAP50 gradually decreases and the FPS begins to decrease. Therefore, this experimental example chooses to connect two DCNv4-Dyhead modules in series in the segmentation model.

[0112] Table 3 Comparison of the impact of the number of DCNv4-Dyhead modules in series on model performance

[0113]

[0114] This experimental example also analyzes the impact of pruning methods on indicators such as model accuracy, network complexity, and inference speed. In view of the complex dynamic characteristics of substation near-electric operation scenarios, the aim is to reduce the computational burden while maximizing model accuracy. Specific data are shown in Table 4. Model ① is the baseline model of YOLOv8s, model ② is the improved YOLOv8s segmentation model after introducing the DCNv4-Dyhead dynamic detection head, and models ③, ④, and ⑤ are pruned versions of the LAMP method with pruning rates of 2.0, 2.5, and 3.0, respectively.

[0115] From the experimental results, it can be seen that with the increase of pruning rate, the mAP50 index of the model shows a downward trend, especially when the speed-up is 2.5 and 3.0, the accuracy loss is more significant. Through the LAMP pruning technology, the model can maintain a high accuracy (mAP50 = 0.831) when the pruning rate Speed-up is 2.0, which is only 0.006 lower than model ②. At the same time, the model weight is reduced to 18.2MB, GFLOPs is reduced to 28.5, and the inference speed is significantly improved to 222.7FPS, showing the obvious optimization of pruning on computing efficiency. In contrast, in the pruned versions with speed-up of 2.5 and 3.0, although the model continues to reduce the consumption of computing resources, the accuracy decreases significantly, especially when the speed-up is 3.0, the mAP50 drops to 0.787, indicating that pruning may have affected the segmentation performance of the model at this time.

[0116] Table 4 Model pruning effect

[0117]

[0118] Combining the experimental results with the actual needs of the substation near-power operation scenario, a pruning rate Speed-up of 2.0 is considered to be an ideal balance point. Therefore, this experimental example uses model ③ as a fine division model for plane operation elements. Figure 5 It can be seen that the number of channels before and after pruning is compared under this pruning rate. The number of channels in each layer after pruning is maintained at a relatively low number. The model can not only cope with the changing operating elements in the scene, but also achieve the best optimization effect in inference speed and computing resource occupancy. It is suitable for deployment on edge computing devices to achieve efficient real-time identification of key elements of near-electric operations.

[0119] Finally, this experimental example compares and analyzes the accuracy and lightweightness of the segmentation model among multiple models, as shown in Table 5. Considering the lightweight accelerated reasoning of the model at the edge, many acceleration libraries and reasoning frameworks currently only support YOLOv5 and YOLOv8, so this experimental example mainly selects these two versions for experimental comparison.

[0120] PR curves between different models are shown in Figure 6According to the data comparison in Table 5, although the anchor-based YOLOv5 performs well on large rigid targets in substations (such as vehicles and transformers), the anchor frame has limited representation capabilities for variable elements such as non-rigid targets (such as workers, telescopic fences, etc.). YOLOv8 is based on the anchor-free design and is used as the benchmark model in this experimental example. It also shows better performance in the fine segmentation of variable and non-rigid targets. Based on this, this experimental example further replaces the DCNv4-Dyhead to improve the model, which can better capture subtle deformations and position changes. The mAP50 reaches 0.831, and the mAP50-95 shows a high generalization ability. The comparison of the fine segmentation effects between different models can be seen in Figure 7 .

[0121] Table 5 Comparison of different models

[0122]

[0123] Ultimately, while balancing indicators such as model size, computational overhead, and inference speed, the model in this paper can effectively adapt to complex and changeable near-electric operation scene detection tasks.

[0124] This experiment uses various visible light images of substation operation scenes and associated binocular images to analyze the performance of the fused model. The binocular vision model used has a depth perception range of 0.1-20m, and the depth information is accurate within this range. The fusion effect of different planar positioning methods can be seen in Figure 8 , the ranging reference value comes from the centimeter-level accuracy data of the laser radar.

[0125] exist Figure 8 In the scene shown, the target detection task uses the center point range mean depth assignment method to measure the distance. Compared with the refined segmentation multi-point assignment method proposed in the present invention, the distance measurement result shows discontinuity, and the singleness of the distance between elements leads to obvious actual distance errors. For example, Figure 8 In scenario (a), the actual minimum distance between the vehicle and the telescopic fence is 2.21m, while the target detection and ranging method outputs 4.52m, and the method of the present invention outputs 2.25m. This shows that the refined segmentation method of the present invention effectively eliminates the redundant background that is compact to the key elements, ensuring that the plane positioning content is concentrated on the key elements, so as to achieve a distance calculation that is closer to the truth.

[0126] according to Figure 8The data of multiple ranging outputs in this experimental example are sorted out in a comparison table, see Table 6. When the ranging is less than 20m, the ranging effect of the method of the present invention is close to the fitting reference value, and the error is kept at the decimeter level. However, the data beyond 20m is mainly affected by the limitations of the distance perception of the binocular vision model. For example, in scene (c), the depth information around the insulator is insufficient. Despite this, the improved model based on DCNv4-Dyhead still shows good planar positioning effect.

[0127] Table 6 Comparison of distance output of dynamic identification model of operational risk

[0128]

[0129] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for dynamic identification of risks of near-electrical operations integrating fine segmentation and binocular vision, characterized in that: The method includes the following steps: S101. Obtain the left view image of substation near-electrification operation captured by a binocular camera; S102. Initialize the segmentation model, binocular three-dimensional ranging model, and calibration parameters of the binocular camera; S103. Identify and extract key operation elements in the left view image through the segmentation model. The segmentation result includes each detected instance and its two-dimensional position; S104. For each instance, predict its three-dimensional position in space through the binocular three-dimensional ranging model to obtain the three-dimensional coordinates of the instance; S105. Calculate the Euclidean distance between different key operation elements according to the three-dimensional coordinates of each instance to obtain the minimum distance between key operation elements; S106. Judge whether there is an operation risk according to the minimum distance between different key operation elements.

2. According to the method of claim 1, the method is characterized by: The segmentation model adopts the YOLOv8 segmentation model.

3. According to claim 2, a method for dynamic identification of risks of near-electrical operations integrating fine segmentation and binocular vision is characterized in that: The DCNv4-Dyhead detection head replaces the baseline model detection head of the YOLOv8 segmentation model. The DCNv4-Dyhead detection head replaces the 3×3 convolution in the Dyhead spatial attention with DCNv4 (DeformableConvNets v4), and removes unnecessary softmax normalization in spatial aggregation. The spatial attention π of the DCNv4-Dyhead detection head S (·)The calculation formula is as follows: In the above formula, the feature tensor F∈R L×S×C , S represents the spatial location information, C represents the channel information, L is the number of input feature layer levels, G is the total number of clustering groups, K is the number of sparse sampling locations, and w l,g represents the weight parameter, which is used to control the weight relationship between the lth channel and the gth group, F(l; p k +Δp gk ; c) indicates that in the lth channel, position p k After an offset of Δp gk The offset is used to calculate the dynamic convolution kernel, c represents the channel dimension, m represents the feature map value after gk represents the modulation scalar of the k-th sampling point in the g-th group, normalized along the K dimension using the softmax function.

4. According to claim 2, a method for dynamic identification of risks of near-electrical operations integrating fine segmentation and binocular vision is characterized in that: Introduce the Layer-Adaptive Magnitude-based Pruning (LAMP) method into the segmentation model. Determine the pruning priority by calculating the LAMP score of each layer's weights to maintain the balance between model performance and sparsity. Specifically, it includes: flatten each weight tensor into a one-dimensional vector, sort it in ascending order according to the given index map. Whenever u < v holds and |W[u]| ≤ |W[v]|, where W[u] represents the u-th element in the weight tensor, the LAMP score for this element is defined as: u represents the index of the weight element currently being calculated, and v represents the index of other weight elements relative to u.

5. According to the method of claim 1, the method is characterized by: For each instance, predict its three-dimensional position in space through the binocular three-dimensional ranging model to obtain the three-dimensional coordinates of the instance, specifically including the steps: S201. Calculate a high-quality disparity map through AANet (Adaptive Aggregation Network) stereo matching, and convert the disparity map into a depth map that associates pixel coordinates (u, v) with Z-axis information. The conversion relationship is: In the above formula, Z is the depth value, f is the camera focal length, b is the baseline distance of the binocular camera, and disp is the disparity value; S202. Combine the depth map and the internal parameters of the binocular module to reconstruct the real point coordinates (X, Y, Z) of the corresponding pixel points in the world coordinate system, so as to correspond the two-dimensional pixel positions in the image to the three-dimensional positions in the actual scene. The calculation formula is: where c x , c y Indicates the location of the principal point; S203. Select multiple points with depth values and spans meeting the requirements for depth assignment within the effective depth range of the same key operation element; S204. Calculate the Euclidean distance between the nearest depth assignment point pairs of two key operation elements and return the minimum distance between key operation elements.

Images