Target space ranging and operation safety warning method based on monocular view geometry

Abstract

The application provides a target space ranging and operation safety early warning method based on single view geometry, obtains multiple scene operation images through a substation monitoring camera and an operation personnel law enforcement instrument, supplements general data set samples to improve diversity, and constructs a substation operation personnel foot key point data set; a key point detector YOLO-P3C is used to obtain two-dimensional information of personnel foot key points, a feature extraction module CIDW, a feature enhancement module CDA and a multi-scale fusion module CEM are integrated to solve the problems of inaccurate small target key point positioning and large model calculation in a complex scene; according to the conversion relationship between the world coordinate system and the pixel coordinate system, the depth information is deduced by using the reverse coordinate conversion formula combined with the prior height, and an error compensation model (LEC) based on reverse coordinate conversion is proposed to correct the inherent nonlinear depth deviation in the monocular reconstruction process, and high-precision three-dimensional reconstruction under monocular vision is realized; finally, the head coordinates are solved based on single view metrology, the safety distance between the personnel and the live region is calculated, and the hierarchical alarm threshold is set, the multi-level early warning function is realized, and reliable support is provided for the safety control of the substation site operation.

Description

Technical Field

[0001] This invention belongs to the field of power equipment safety monitoring technology, and relates to a target space ranging and operation safety early warning method based on single-view geometry. Background Technology

[0002] As a core component of the power system, substations undertake the voltage conversion task for long-distance power transmission, playing a vital role in ensuring power quality and safe equipment operation. However, the high-voltage environment and complex operating scenarios pose severe challenges to the safety management of substations.

[0003] In response to such safety incidents, the power industry has established relevant standards. For example, the Chinese national standard GB 26860-2011, "Electrical Safety Work Regulations for Power Plants and Substations," specifies safe distances for equipment without power outages (e.g., 5m for 500kV). Substations also reduce accident risks through measures such as setting up warning signs and strengthening safety patrols. While these solutions can reduce the occurrence of safety incidents to some extent, they rely heavily on the subjective awareness of maintenance personnel, which introduces significant uncertainty. Furthermore, prolonged work hours can lead to decreased attention spans, and workers may overlook safety distance regulations due to negligence, fatigue, or a lack of sufficient safety awareness. Relying solely on subjective awareness and existing passive preventative measures is far from sufficient; effective active monitoring methods for safety distance monitoring are urgently needed. Therefore, a target space ranging and operational safety early warning method based on single-view geometry needs to be designed to address these issues. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a target space ranging and operation safety early warning method based on single-view geometry. The method obtains the intrinsic and extrinsic parameters and distortion coefficients of the camera under a fixed viewpoint through the Zhang Zhengyou calibration method. Combined with prior height information and inverse coordinate transformation formula, the two-dimensional pixel coordinates are converted into three-dimensional world coordinates. The LEC model is designed to correct the three-dimensional coordinates, solve the nonlinear error of distant pixels, and significantly enhance the robustness and ranging accuracy of the method. Finally, the Euclidean distance formula is used to calculate the distance between a person and multiple energized areas and parts. Based on the safety distance specifications, customizable multi-level safety distance early warning is realized, and the automatic monitoring, ranging and early warning of substation safety distances are completed.

[0005] To achieve the above-mentioned technical effects, the technical solution adopted by the present invention is as follows: A target spatial ranging and operation safety early warning method based on single-view geometry includes the following steps: S1. Acquire images of operations in multiple scenarios and supplement them with COCO dataset samples to improve diversity; perform foot key point annotation and data cleaning on all images to construct a dataset of foot key points for substation workers. S2, construct the YOLO-P3C keypoint detector. The detector integrates a multi-scale pyramid structure of feature extraction module CIDW, feature enhancement module CDA, and fusion CEM module to detect foot keypoints; S3, based on the transformation relationship between the world coordinate system, camera coordinate system, image coordinate system and pixel coordinate system, uses the inverse coordinate transformation formula and combines the prior height value to map the two-dimensional image coordinates to the three-dimensional world coordinates and derive the depth information. S4 proposes an error compensation model based on reverse coordinate transformation, LEC, which uses the camera coordinate system depth and the initial world coordinates as geometric priors to construct an explicit compensation function to correct the nonlinear depth deviation in the monocular reconstruction process and obtain the corrected three-dimensional coordinates. S5, based on the corrected three-dimensional coordinates, uses a single-view metrology method to solve the key points of the head, calculates the safe distance between maintenance personnel and live areas and parts, and realizes hierarchical early warning and visualization display according to preset thresholds.

[0006] Preferably, in step S1, six key points are selected in the dataset: the right big toe, the right little toe, the right heel, the left big toe, the left little toe, and the left heel; low-resolution, motion-blurred, and severely occluded samples are removed during data cleaning.

[0007] Preferably, in step S2, the feature extraction module CIDW is specifically implemented as follows: The feature extraction module CIDW is constructed, which draws on the design of multi-branch receptive field, partial convolution strategy and multi-dimensional attention mechanism. The conventional two-dimensional convolution in the C3k structure is replaced with a partial convolution module, which only processes one-quarter of the input channels. The multi-dimensional attention mechanism MCA is introduced to apply attention in the three dimensions of height, width and channels respectively, which enhances the ability to capture fine-grained features of small targets. Preferably, in step S2, the feature enhancement module CDA is specifically implemented as follows: A feature enhancement module CDA is constructed, which introduces the idea of ​​deformable sampling into the self-attention mechanism and combines the CSP structure with the feedforward network to achieve feature fusion. The offset corresponding to each query point is predicted through a lightweight convolutional network, and the sampling position is dynamically adjusted so that the attention distribution can adaptively focus on the dense foot key point area according to the content. Preferably, in step S2, the integration of the CEM module into the multi-scale pyramid structure is specifically implemented as follows: The CEM module is integrated into a multi-scale pyramid structure. High-resolution semantic feature representation is enhanced by channel-space joint attention and multi-receptive field convolution aggregation. The statistical properties of the input features are extracted using global pooling to generate channel attention weights, spatial statistical properties are calculated to generate spatial attention weights, and depth-separable convolutions of different sizes are used to expand channel features and fuse information from different receptive fields.

[0008] Preferably, in step S3, 3D reconstruction is achieved by utilizing constant parameters of the camera at a fixed angle, combined with prior height values ​​and the inverse coordinate transformation formula; the specific transformation is as follows: Establish the transformation formula between world coordinates and pixel coordinates: ; in, For depth information; u , v () represents the pixel coordinates of a point on the image plane; X w , Y w , Z w () represents the coordinates of a point in the world coordinate system; R For rotation matrix, T It is a translation vector; Based on the inverse coordinate transformation relationship, the known target height value is introduced. The depth information of the target is derived using the following formula. : ; ; in, , P (2, 0) represents the value in the third row and first column of the matrix; based on the calculated... It can derive three-dimensional world coordinates from two-dimensional pixel coordinates and use the derived three-dimensional world coordinates as geometric prior inputs for subsequent LEC models.

[0009] Preferably, in step S4, the error compensation model LEC based on reverse coordinate transformation is as follows: Depth of camera coordinate system with initial world coordinates As a geometric prior, an explicit compensation function for long-range errors is constructed; Multinomial regression is used to jointly compensate the error term, with the depth in the camera coordinate system as the basis. and horizontal coordinates As a joint input, a polynomial basis function is constructed to capture the nonlinear coupling of errors in the depth and lateral directions: ; ; in, X e and Y e These are the error compensation values ​​in the X-axis and Y-axis directions, respectively, with coefficients { a i}and{ b i The interaction term was obtained by least squares fitting. Z c X w and quadratic terms Z c 2 , X w 2 Used to capture the nonlinear coupling between lateral coordinates and depth variations; higher-order terms Z c 3 Used to model the trend of depth error increasing with the square or even cube of distance.

[0010] Preferably, in step S5, the specific method for implementing safe distance monitoring and early warning is as follows: Obtain the target's 3D coordinates: For charged areas, the LEC model is used to correct coordinates; for maintenance personnel, the head coordinates are solved by combining foot key point coordinate mapping and single-view metrology methods, and the mapping relationship between pixel coordinates and physical height is established using the following formula: ; in, η Indicates the scaling factor. p b and p t These represent the coordinates of the operations and maintenance personnel at the bottom and top of the pixel coordinate system, respectively. θ and μ These represent the horizontal cancellation line and the vertical cancellation point in the pixel coordinate system, respectively. Calculate distance and issue a warning: Based on the known world coordinates of the charged parts and the solved personnel coordinates, the safe distance is calculated using the following formula. d : ; Based on the preset graded early warning strategy, different distance thresholds are set. When the monitoring distance is lower than the corresponding threshold, the corresponding color signal and visual warning are triggered.

[0011] Preferably, a computer device includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the target spatial ranging and operation safety early warning method based on single-view geometry.

[0012] Preferably, the computer-readable storage medium stores computer instructions for causing the computer to execute the target space ranging and operation safety early warning method based on single-view geometry.

[0013] The beneficial effects of this invention are as follows: This invention provides a target spatial ranging and operational safety early warning method based on single-view geometry. It achieves efficient ranging through a keypoint detector, inverse coordinate transformation formula, and LEC model. Drawing on multi-branch receptive field design, partial convolution strategy, and multi-dimensional attention mechanism, a feature extraction module (CIDW) with multi-scale representation capability and high computational efficiency is constructed, effectively enhancing the capture capability of fine-grained features of small targets. A CDA module is proposed, introducing deformable sampling into the self-attention mechanism and combining it with a CSP structure and feedforward network to achieve efficient feature fusion. This allows the attention distribution to adaptively focus on dense foot keypoint regions according to content, thereby improving detection accuracy under complex backgrounds and occlusion conditions. Inspired by multi-scale fusion strategies, a CEM module is proposed, which enhances high-resolution semantic feature expression through channel-space joint attention and multi-receptive field convolution aggregation, effectively improving the localization accuracy of small-scale keypoints. An LEC model is proposed, which uses the camera coordinate system depth Z... c The initial world coordinates Z obtained based on the inverse transformation w As a geometric prior, an explicit compensation function for long-range errors is constructed to correct the inherent nonlinear bias in monocular reconstruction. Compared with traditional binocular vision and LiDAR solutions, this method is lower in cost and easier to deploy, and can make full use of existing monocular camera resources in substations, providing a cost-effective real-time safety monitoring solution for the power IoT sensing layer. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating an embodiment of the present invention; Figure 2 This is a statistical distribution diagram of the effective pixel area of ​​foot key points in the dataset of this embodiment of the invention; Figure 3 This is a structural diagram of the YOLO-P3C model according to an embodiment of the present invention; Figure 4 This is a structural diagram of the C3k2 module according to an embodiment of the present invention; Figure 5 This is a structural diagram of the CIDW module in an embodiment of the present invention (cidw=False). Figure 6 This is a structural diagram of the CIDW module in an embodiment of the present invention (cidw=True). Figure 7 This is a structural diagram of the CDA module according to an embodiment of the present invention; Figure 8 This is a structural diagram of the CEM module according to an embodiment of the present invention; Figure 9 This is a convergence curve of the training loss function in an embodiment of the present invention; Figure 10 This is a performance comparison chart of the models in this embodiment of the invention; Figure 11 This is a diagram showing the relationship between the four coordinate systems in an embodiment of the present invention; Figure 12 This is a distance measurement error analysis diagram of a multi-layered terrain scene monitoring video in an embodiment of the present invention; Figure 13 This is a method verification diagram according to an embodiment of the present invention; Figure 14 This is a case verification diagram for an embodiment of the present invention. Detailed Implementation

[0015] Example 1: A target spatial ranging and operation safety early warning method based on single-view geometry includes the following steps: S1. Acquire images of operations in multiple scenarios and supplement them with COCO dataset samples to improve diversity; perform foot key point annotation and data cleaning on all images to construct a dataset of foot key points for substation workers. S2, construct the YOLO-P3C keypoint detector. The detector integrates a multi-scale pyramid structure of feature extraction module CIDW, feature enhancement module CDA, and fusion CEM module to detect foot keypoints; S3, based on the transformation relationship between the world coordinate system, camera coordinate system, image coordinate system and pixel coordinate system, uses the inverse coordinate transformation formula and combines the prior height value to map the two-dimensional image coordinates to the three-dimensional world coordinates and derive the depth information. S4 proposes an error compensation model based on reverse coordinate transformation, LEC, which uses the camera coordinate system depth and the initial world coordinates as geometric priors to construct an explicit compensation function to correct the nonlinear depth deviation in the monocular reconstruction process and obtain the corrected three-dimensional coordinates. S5, based on the corrected three-dimensional coordinates, uses a single-view metrology method to solve the key points of the head, calculates the safe distance between maintenance personnel and live areas and parts, and realizes hierarchical early warning and visualization display according to preset thresholds.

[0016] Preferably, in step S1, six key points are selected in the dataset: the right big toe, the right little toe, the right heel, the left big toe, the left little toe, and the left heel; low-resolution, motion-blurred, and severely occluded samples are removed during data cleaning.

[0017] Preferably, in step S2, the feature extraction module CIDW is specifically implemented as follows: The feature extraction module CIDW is constructed, which draws on the design of multi-branch receptive field, partial convolution strategy and multi-dimensional attention mechanism. The conventional two-dimensional convolution in the C3k structure is replaced with a partial convolution module, which only processes one-quarter of the input channels. The multi-dimensional attention mechanism MCA is introduced to apply attention in the three dimensions of height, width and channels respectively, which enhances the ability to capture fine-grained features of small targets. Preferably, in step S2, the feature enhancement module CDA is specifically implemented as follows: A feature enhancement module CDA is constructed, which introduces the idea of ​​deformable sampling into the self-attention mechanism and combines the CSP structure with the feedforward network to achieve feature fusion. The offset corresponding to each query point is predicted through a lightweight convolutional network, and the sampling position is dynamically adjusted so that the attention distribution can adaptively focus on the dense foot key point area according to the content. Preferably, in step S2, the integration of the CEM module into the multi-scale pyramid structure is specifically implemented as follows: The CEM module is integrated into a multi-scale pyramid structure. High-resolution semantic feature representation is enhanced by channel-space joint attention and multi-receptive field convolution aggregation. The statistical properties of the input features are extracted using global pooling to generate channel attention weights, spatial statistical properties are calculated to generate spatial attention weights, and depth-separable convolutions of different sizes are used to expand channel features and fuse information from different receptive fields.

[0018] Preferably, in step S3, 3D reconstruction is achieved by utilizing constant parameters of the camera at a fixed angle, combined with prior height values ​​and the inverse coordinate transformation formula; the specific transformation is as follows: Establish the transformation formula between world coordinates and pixel coordinates: ; in, For depth information; u , v () represents the pixel coordinates of a point on the image plane; X w , Y w , Z w () represents the coordinates of a point in the world coordinate system; R For rotation matrix, T It is a translation vector; Based on the inverse coordinate transformation relationship, the known target height value is introduced. The depth information of the target is derived using the following formula. : ; ; in, , P (2, 0) represents the value in the third row and first column of the matrix; based on the calculated... It can derive three-dimensional world coordinates from two-dimensional pixel coordinates and use the derived three-dimensional world coordinates as geometric prior inputs for subsequent LEC models.

[0019] Preferably, in step S4, the error compensation model LEC based on reverse coordinate transformation is as follows: Depth of camera coordinate system with initial world coordinates As a geometric prior, an explicit compensation function for long-range errors is constructed; Multinomial regression is used to jointly compensate the error term, with the depth in the camera coordinate system as the basis. and horizontal coordinates As a joint input, a polynomial basis function is constructed to capture the nonlinear coupling of errors in the depth and lateral directions: ; ; in, X e and Y e These are the error compensation values ​​in the X-axis and Y-axis directions, respectively, with coefficients { a i}and{ b i The interaction term was obtained by least squares fitting. Z c X w and quadratic terms Z c 2 , X w 2 Used to capture the nonlinear coupling between lateral coordinates and depth variations; higher-order terms Z c 3 Used to model the trend of depth error increasing with the square or even cube of distance.

[0020] Preferably, in step S5, the specific method for implementing safe distance monitoring and early warning is as follows: Obtain the target's 3D coordinates: For charged areas, the LEC model is used to correct coordinates; for maintenance personnel, the head coordinates are solved by combining foot key point coordinate mapping and single-view metrology methods, and the mapping relationship between pixel coordinates and physical height is established using the following formula: ; in, η Indicates the scaling factor. p b and p t These represent the coordinates of the operations and maintenance personnel at the bottom and top of the pixel coordinate system, respectively. θ and μ These represent the horizontal cancellation line and the vertical cancellation point in the pixel coordinate system, respectively. Calculate distance and issue a warning: Based on the known world coordinates of the charged parts and the solved personnel coordinates, the safe distance is calculated using the following formula. d : ; Based on the preset graded early warning strategy, different distance thresholds are set. When the monitoring distance is lower than the corresponding threshold, the corresponding color signal and visual warning are triggered.

[0021] Preferably, a computer device includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the target spatial ranging and operation safety early warning method based on single-view geometry.

[0022] Preferably, the computer-readable storage medium stores computer instructions for causing the computer to execute the target space ranging and operation safety early warning method based on single-view geometry.

[0023] Example 2: This embodiment provides a process for applying this method in a practical application, as detailed below: like Figure 1 A target spatial ranging and operation safety early warning method based on single-view geometry is disclosed, which includes the following steps: S1. Based on the substation monitoring camera and the law enforcement recorder of the maintenance personnel, obtain multi-scene operation images, supplement the COCO dataset samples to improve diversity, and perform foot key point annotation and data cleaning on all images to construct a substation operator foot key point dataset. S2. Construct the YOLO-P3C keypoint detector, which integrates the feature extraction module CIDW, the feature enhancement module CDA, and the multi-scale pyramid structure of the fused CEM module to achieve high-precision and lightweight detection of foot keypoints; S3. Based on the transformation relationship between the world coordinate system, camera coordinate system, image coordinate system and pixel coordinate system, the inverse coordinate transformation formula is used and combined with the prior height value to map the two-dimensional image coordinates to three-dimensional world coordinates and derive the depth information. S4. An error compensation model (LEC) based on reverse coordinate transformation is proposed. The camera coordinate system depth and the initial world coordinates are used as geometric priors to construct an explicit compensation function to correct the nonlinear depth deviation in the monocular reconstruction process. S5. Based on the corrected three-dimensional coordinates, the key points of the head are solved using the single-view metrology method, the safe distance between maintenance personnel and live areas and parts is calculated, and graded early warning and visualization are realized according to the preset threshold.

[0024] In the preferred scheme, in S1, a dedicated dataset containing 3076 images was constructed based on substation monitoring cameras and enforcement recorders used by maintenance personnel. Among these, 1946 images depict substation scenes, collected at multiple substation sites under varying lighting conditions, camera angles, and work postures. An additional 1130 human images covering various scenes were randomly selected from the COCO dataset for supplementary annotation to enhance feature diversity. Of all the images, 734 showed partial occlusion of key foot points. Since nighttime operations at substations are relatively rare, all images were collected during the day, with 132 images taken on cloudy days and the remainder in well-lit conditions. Six key points were selected for the dataset: right big toe, right little toe, right heel, left big toe, left little toe, and left heel. To ensure data quality, low-resolution, motion-blurred, or severely occluded samples were removed. This dataset covers both near-field and far-field work scenarios, providing support for performance evaluation of the proposed depth correction mechanism.

[0025] Preferably, the scale distribution of foot keypoints in the dataset was analyzed to objectively assess the challenges of the proposed task. Since keypoints are defined in coordinate form and do not possess size information, a proportional substitution method was used to estimate the effective pixel area of ​​the foot keypoints, set to 2% of the human body bounding box area. This estimation result can characterize the local feature regions required for accurate localization.

[0026] Figure 2 The final distribution is shown; statistical data indicates that, according to the definition of the MS COCO dataset, 33.17% of the targets are "small targets" (area < 32² pixels), and 34.20% are "medium targets" (area 32² ≤ area < 96² pixels). The high proportion of small-to-medium targets (> 67%) and the significant scale difference highlight the inherent difficulty of this task and also validate the necessity of improving the proposed model.

[0027] like Figure 3As shown in S2, a new keypoint detector, YOLO-P3C, is constructed based on the YOLOv11n-pose model. It consists of three parts: a feature extraction module (CIDW), a feature enhancement module (CDA), and a multi-scale pyramid structure integrating a CEM module. These three components complement each other, forming a complete optimization path from low-level feature extraction to high-level semantic fusion, significantly improving the detection accuracy and robustness of foot keypoints.

[0028] Preferably, by drawing on multi-branch receptive field design, partial convolution strategy and multi-dimensional attention mechanism, a feature extraction module CIDW is constructed that has both multi-scale expressive ability and high computational efficiency, effectively enhancing the ability to capture fine-grained features of small targets.

[0029] like Figure 4 As shown, the C3k2 module in the baseline model inherits from the C2f structure, and the selection of the C3k or Bottleneck structure is controlled by the parameter c3k (True or False).

[0030] The C3k2 module performs well in large target feature extraction, but its performance in this task is limited because it lacks scale optimization for small targets, while small target detection usually requires stronger multi-scale perception capabilities.

[0031] like Figure 5 As shown, when c3k is False, the two fixed 3×3 convolutional kernels in Bottleneck are decomposed into multiple groups of smaller convolutional kernels: 1 / 3 of the channels use 3×3 convolutional kernels, 1 / 3 of the channels use 1×11 convolutional kernels, and the remaining 1 / 3 of the channels use 11×1 convolutional kernels. Due to the use of depthwise separable convolution, the computational cost is reduced, and the inference speed is improved. At the same time, a residual structure is introduced to preserve the original features and reduce information loss.

[0032] The original module achieves high lightweight performance through the c3k design. When c3k is True, the C3k2 module uses the C3k structure, convolving only half of the input channels. To further improve lightweight performance, the two regular 2D convolutions in the C3k structure are replaced with a partial convolution module, processing only one-quarter of the input channels. To compensate for the decrease in feature extraction capability caused by the reduction in convolutions, a multi-dimensional attention mechanism (MCA) is added. This attention mechanism applies attention in the height (H), width (W), and channel (C) dimensions respectively, enhancing feature representation through average pooling, standard deviation pooling, and one-dimensional convolution. It can dynamically adjust the convolution kernel according to the number of channels, optimizing computational efficiency and detection performance, and enhancing the detection capability of small objects, such as... Figure 6 As shown.

[0033] Preferably, a CDA module is proposed, which introduces the idea of ​​deformable sampling into the self-attention mechanism and combines the CSP structure with the feedforward network to achieve efficient feature fusion, so that the attention distribution can adaptively focus on the dense foot key point area according to the content, thereby improving the detection accuracy under complex background and occlusion conditions.

[0034] In the baseline model, the C2PSA module combines a CSP structure with a PSA multi-head self-attention mechanism, processing the input feature map by applying convolutional kernels of different sizes in parallel. Each convolutional kernel extracts features within a different receptive field, effectively capturing multi-scale spatial information. However, because it relies on fixed-size convolutional kernels to process the global feature map in parallel, its computational complexity is high, especially at high input resolutions, potentially leading to excessive consumption of computational resources. Furthermore, the fixed-field convolutional operation lacks dynamic adaptability to target location, which may make it difficult to accurately focus on key areas when detecting small and densely packed foot points in substations, thus affecting detection accuracy and efficiency.

[0035] Drawing inspiration from the attention mechanism DAttention, its core principle lies in combining the idea of ​​deformable convolution. It uses a lightweight convolutional network to predict the offset corresponding to each query point, thereby dynamically adjusting the sampling positions of the key and value. This mechanism breaks through the limitations of a fixed grid, adaptively focusing on important regions such as foot keypoints based on the input feature content. Compared to the fixed multi-scale convolution of C2PSA, it significantly improves the ability to capture details of small objects in complex scenes. Furthermore, DAttention supports multiple positional encoding methods, enhancing the model's ability to model spatial relationships. Based on this, a CDA module is proposed, such as... Figure 7 As shown, it integrates the DAttention self-attention mechanism, CSP structure and feed-forward network FFN, taking into account the advantages of both.

[0036] Preferably, inspired by multi-scale fusion strategies, a CEM module is proposed, which enhances high-resolution semantic feature representation through channel-space joint attention and multi-receptive field convolution aggregation, thereby effectively improving the localization accuracy of small-scale key points.

[0037] The baseline model employs an FPN-PAN design for the neck structure, improving the utilization of high-level semantics and low-level localization information. However, the FPN-PAN design still has limitations in foot keypoint detection. Foot keypoints, as small targets, rely on high-resolution feature maps (P3 / 8, stride=8) for prediction. While the original multi-scale design optimizes overall detection performance, it prioritizes balancing large, medium, and small targets, paying insufficient attention to high-resolution features, thus affecting the accuracy of foot keypoint localization.

[0038] To address this, the CEM module is proposed. Compared to traditional attention mechanisms that focus only on a single channel or spatial dimension, CEM achieves refined enhancement of small target features through a dynamic feature selection mechanism that combines channel and spatial dimensions.

[0039] like Figure 8 As shown, CEM utilizes the statistical properties of global pooling to extract input features, generating channel attention weights to enhance semantic information related to foot keypoints while suppressing interference from irrelevant channels. Based on weighted features, CEM calculates spatial statistical properties to generate spatial attention weights, dynamically focusing on key foot regions and enhancing local contextual representation. To fully utilize multi-scale information, CEM integrates the multi-scale convolution strategy of the CIDW module, employing 1×11, 11×1, and 3×3 depthwise separable convolutions to expand channel features and fuse information from different receptive fields, enriching feature representation. To balance performance and lightweight design, CEM borrows from the EMCAD attention mechanism, combining adaptive maximum pooling (AMP) and adaptive average pooling (AAP) to obtain global channel information, capturing local spatial features through channel-dimensional maximum and average pooling.

[0040] Preferably, the degree of impact of different improvement modules on model performance is shown in Table 1 below: Table 1: Ablation experiment results;

[0041] Among them, FLOPs (Floating Point Operations per Second) represents the total number of floating point operations of the neural network model, and Params (Parameters) represents the total number of learnable parameters in the model. These two metrics are used to measure model complexity, and lower values ​​are better. Average Precision (AP) is used to evaluate the detection accuracy of the model, and Frames per Second (FPS) represents the number of frames processed by the model in one second, used to measure the real-time performance of the model. Higher values ​​for both metrics are better. The calculation of AP depends on the Intersection of Union (IoU) threshold. For keypoint detection models, Object Keypoint Similarity (OKS) is equivalent to IoU, and its calculation formula is as follows: ; in: K This represents the total number of key points on the human target; in this task, there are 6. d kRefers to the first k The Euclidean distance between each predicted keypoint and the actual keypoint measures the degree of deviation between the predicted keypoint location and the actual keypoint location. s The scale representing the size of the human body target. σ k It is the first k The standard deviation of a key point is used to measure the uncertainty or allowable error range of the key point's location. m k It is a binary mask value used to mark the first k Is the first key point visible? k The key points are visible in the image. m k =1; if invisible (e.g., obscured), then m k =0. Specifically, AP represents OKS Detection accuracy when the threshold is 0.5~0.95. AP 50 express OKS Detection accuracy at a threshold of 0.5. All comparisons were conducted under the same preset conditions to ensure the accuracy and impartiality of the experiment.

[0042] The results show that the CDA and CEM modules slightly improve detection accuracy without significantly increasing the number of model parameters and complexity. The combined application of both modules significantly improves detection accuracy. The CIDW module reduces model complexity and the number of parameters through depth-separable multi-scale convolution design and improved partial convolution design, while slightly enhancing detection performance. Ultimately, the proposed model structure optimization improves AP to 0.773, an increase of 6.32%. This improvement in detection accuracy means the model can more accurately locate foot key points, thereby reducing misjudgments of safe distances due to key point offsets. Furthermore, the proposed model requires only 6.2 GFLOPs and 2.34 M parameters while maintaining high detection accuracy, exhibiting excellent lightweight characteristics. It can be stably deployed on substation monitoring terminals or edge embedded devices, supporting long-term low-power operation. In addition, the model inference speed reaches 243 f / s, further enhancing the system's real-time performance compared to the baseline model (222 f / s), enabling timely safety warnings when the distance between workers and live equipment changes drastically.

[0043] Preferably, multi-dimensional comparative experiments were designed to comprehensively verify the advancement and effectiveness of the proposed method; Table 2 shows the experimental results: Table 2: Comparison results of different models in the embodiments of the present invention;

[0044] Table 2 shows that YOLOv12 performs comparably to the baseline models, and is comprehensively superior to YOLOv8. EfficientVIT, RepViT, and ConvNeXt V2 have slightly higher accuracy than the YOLO series, but their computational complexity increases significantly. In comparison with dedicated attitude estimation models, RTMO shows strong competitiveness (AP 0.761), but its parameter count (9.90 M) and computational cost (8.9 GFLOPs) are relatively large, with an FPS of only 118 f / s, making it difficult to meet the high-frequency real-time requirements of edge computing. In contrast, YOLO-P3C achieves the highest accuracy while maintaining the lowest computational cost and the fastest inference speed, significantly outperforming dedicated models such as CID and DEKR in overall performance. In the generalization verification experiments, each module demonstrated good transferability. After introducing the CIDW module, the GFLOPs of Hyper-YOLO and YOLO-Pose decreased significantly by 40% and 35%, respectively, demonstrating CIDW's significant advantage in reducing computational redundancy. Although AP fluctuated slightly on some models, it greatly improved inference efficiency. After integrating the CDA and CEM modules, the AP values ​​were steadily improved compared to the baseline models, verifying the universality and effectiveness of the proposed modules under different architectures.

[0045] Preferably, the changes in the loss function during the YOLO-P3C training process were visualized and analyzed to verify the stability and convergence characteristics of the proposed model during the training phase.

[0046] Figure 9 The convergence trajectories of the bounding box regression loss (box_loss, dfl_loss), keypoint localization loss (pose_loss, kobj_loss), and classification loss (cls_loss) are shown. Thanks to the improved network structure, each loss term decreases rapidly in the early stages of training, then enters a stable optimization phase. Throughout the 300 training epochs, the curves maintain a smooth and continuously converging trend, without drastic oscillations, gradient explosion, or overfitting, indicating that the model can effectively learn the target features. In particular, towards the end of training (around epoch 290), the overall loss exhibits a step-like decrease. This is because the Mosaic data augmentation strategy is disabled at this stage, allowing the model to perform a final, refined fit on data closer to the true distribution, thereby further improving the convergence quality.

[0047] Preferably, a subset of detected images is selected for comparative analysis to visually demonstrate the model improvement effect.

[0048] like Figure 10As shown, subfigures (a) and (c) illustrate the detection results of the original model, while subfigures (b) and (d) illustrate the detection results of the improved model. The results show that the original model can only roughly detect the human body contour, with significant shifts in foot keypoints. The proposed model, however, can accurately detect foot keypoints, performing well even in scenes with some occlusion. This further validates the effectiveness of the improved method.

[0049] In the preferred embodiment, in S3, constant intrinsic and extrinsic parameters and distortion coefficients of the camera at a fixed angle are used, combined with prior height values ​​and inverse coordinate transformation formulas to achieve three-dimensional reconstruction, mapping the two-dimensional image coordinates to three-dimensional world coordinates.

[0050] The fields of image processing and stereo vision involve four coordinate systems: world coordinate system, camera coordinate system, image coordinate system and pixel coordinate system. The core of these systems is to achieve mutual conversion between pixel coordinates and world coordinates through matrix operations.

[0051] like Figure 11 The diagram illustrates the relationships between four coordinate systems and their transformation mechanisms. Among them: O w - X w Y w Z w It is a world coordinate system used to describe the position of the target object in the real world. O c - X c Y c Z c The camera coordinate system is a coordinate system established on the camera, describing the position of objects from the camera's perspective. o - xy The image coordinate system is used to describe the projection and transmission relationship of an object from the camera coordinate system to the image coordinate system during the imaging process. u - v Pixel coordinates are used to describe the coordinates of image points on a digital image after an object is imaged.

[0052] Based on the relationship between the four coordinate systems, the conversion formula between world coordinates and pixel coordinates can be obtained, as follows: ; in: Z c This represents depth information, specifically the coordinates of a point in the camera coordinate system along the Z-axis (optical axis). u , v () represents the pixel coordinates of a point on the image plane.X w , Y w , Z w ) represents the coordinates of a point in the world coordinate system. R For rotation matrix, T These are translation vectors, used to describe the rotation and translation relationships from the world coordinate system to the camera coordinate system.

[0053] Using the above transformation formula, high-dimensional world coordinates can be mapped to low-dimensional pixel coordinates. However, due to the lack of depth information... Z c Two-dimensional pixel coordinates cannot be directly mapped back to world coordinates. The above formula can be transformed into the following: ; Taking the third term from both sides of the formula and rearranging it, we get: ; in, P 1= R -1 T , P 2= R -1 K -1 [ UV 1] T , P (2, 0) represents the value in the third row and first column of the matrix. The camera intrinsic parameters are shown in the formula. K Rotation matrix R Translation vector T This can be obtained using calibration methods. Therefore, only the target height value needs to be provided. Z w The depth information of the target can then be derived using the inverse coordinate transformation formula. Z c In this task, the foot is used as the key detection target. Considering the approximately 5cm height difference between the foot plane and the ground, the height value is... Z w This can be considered a known constant. Ultimately, based on the calculated... Z c Three-dimensional world coordinates can be derived from two-dimensional pixel coordinates.

[0054] Based on the aforementioned inverse coordinate transformation relationship, the depth information in the camera coordinate system can be... Z c With the predicted initial world coordinates Z w It serves as the geometric prior input to the LEC model, used for subsequent error compensation and nonlinear correction.

[0055] In the preferred scheme, in S4, the LEC model is proposed to define the depth of the camera coordinate system. Z c Compared with the initial world coordinates obtained based on the inverse transformation Z w As a geometric prior, an explicit compensation function for long-range errors is constructed to correct the inherent nonlinear bias in monocular reconstruction.

[0056] In monocular vision ranging, the backprojection process depends on depth. Z c Small deviations in depth estimation can be nonlinearly amplified at long distances, leading to a significant accumulation of ranging errors. To mitigate this problem of nonlinear depth deviation at long distances, an error compensation model based on inverse coordinate transformation (LEC model) is proposed, which systematically corrects depth-related errors through statistical modeling.

[0057] ; As shown in the above formula, pixel coordinates ( u , v ) to world coordinates ( X w , Y w , Z w The back-projection relationship shows that the magnification ratio of the predicted coordinates is proportional to the depth. As the magnification increases, the depth deviation will be amplified into a larger physical distance error, causing depth instability and error accumulation in distant scenes. Therefore, this scheme theoretically analyzes the nonlinear relationship between error and depth, revealing that it mainly exhibits a quadratic growth trend.

[0058] Preferably, to characterize the variation of this error with depth, this paper collected multiple sets of 2D-3D correspondence data from near field to far field and statistically analyzed the residual distribution between predicted coordinates and true coordinates. The results show that the 3D prediction error is related to depth. Z c It exhibits a significant nonlinear correlation and is affected by the predicted coordinates. X w The influence of depth magnification is that pixels at different lateral positions exhibit inconsistency errors. Based on these observations, this paper constructs an LEC model and uses multinomial regression to jointly compensate for the error term.

[0059] Specifically, the LEC model uses depth Z c and horizontal coordinates X wAs a joint input, a polynomial basis function is constructed to capture the nonlinear coupling of errors in the depth and lateral directions.

[0060] ; ; in: X e and Y e These are the error compensation values ​​in the X-axis and Y-axis directions, respectively, with coefficients { a i}and{ b i The interaction term was obtained by least squares fitting. Z c X w and quadratic terms Z c 2 , X w 2 Used to capture the nonlinear coupling between lateral coordinates and depth variations; higher-order terms Z c 3 Used to model the trend of depth error increasing with the square or even cube of distance.

[0061] Preferably, to verify the effectiveness of the model, fitting experiments were conducted on 74 pairs of 2D-3D data, covering the depth range from near field to far field, such as... Figure 12 As shown.

[0062] The goodness of fit is determined by the coefficient of determination. R 2 An evaluation was conducted, and the results showed R 2 A value close to 1 indicates that the polynomial regression can capture the nonlinearity of the error distribution well, effectively correct the prediction bias caused by depth changes, and thus improve the accuracy and robustness of 3D coordinate recovery in distant scenes.

[0063] In summary, the LEC model, by introducing joint constraints on depth and lateral coordinates, realizes a depth error correction mechanism driven by geometric priors, providing theoretical support for improving distance prediction accuracy in subsequent experiments.

[0064] In the preferred embodiment, in S5, based on the output of the key point detector, the inverse coordinate transformation formula and the high-precision three-dimensional reconstruction of the LEC model, the monitoring and early warning of the safe distance between the energized area and energized parts of the substation and personnel are realized.

[0065] The energized area typically refers to the area immediately surrounding the substation facilities, which can be assumed to be a rectangular area with a height value of [missing information]. Z w =0. Based on the inverse coordinate transformation formula, the three-dimensional world coordinates of all points in this region can be derived, and the LEC model can be applied to correct the coordinates to improve accuracy. The two-dimensional image coordinates of key points on the human foot obtained using a keypoint detection algorithm can also be mapped to the corresponding three-dimensional world coordinates. x p , y p , z p The world coordinates of the head keypoints can be derived from the x-axis and y-axis coordinates based on the foot keypoints, while the z-axis coordinates can be calculated using a method based on single-view metrology. This method utilizes the principle of cross-ratio invariance in projective geometry, combined with the vertical vanishing point and horizontal vanishing line (horizon) in the image, to establish a geometric mapping relationship between pixel coordinates and physical height (as shown in the following formula). η Indicates the scaling factor. p b and p t These represent the coordinates of the operations and maintenance personnel at the bottom and top of the pixel coordinate system, respectively. θ and μ These represent the horizontal vanishing line and the vertical vanishing point in the pixel coordinate system, respectively.

[0066] ; For a substation facility with a known equipment model, the world coordinates of its energized parts can be obtained through non-contact methods. x e , y e , z e Finally, the distance between the person and the charged area and parts is calculated using the following formula. d .

[0067] ; Preferably, to verify the effectiveness of the method, an experiment is conducted in a laboratory setting, and the test results are as follows: Figure 13 As shown.

[0068] In this diagram, multiple doors represent live equipment, labeled as targets a, b, and c. The bottom of each door is marked with a red circle and connecting line indicating the live area, while the top of each door is marked with a red rectangle indicating the live part. Detailed detection data is provided, including the actual distance between people and the live areas and parts. d region and d partAfter applying the LEC model, the predicted distance is... d p-region and d p-part Divided into pre-correction distance d pb-region , d pb-part and corrected distance d pa-region , d pa-part Experiments show that in short-range ranging (less than 5 m), the ranging accuracy of this method reaches the millimeter level. In long-range ranging (approximately 10 m), without using the LEC model, the ranging error is as high as 35 cm; after applying the LEC model, the error is reduced to less than 5 cm.

[0069] Preferably, to verify the effectiveness and generalization of the proposed method, cameras were deployed in a substation setting to randomly collect images of personnel for testing. To ensure the accuracy of the evaluation benchmark, a high-precision handheld laser rangefinder (model: T-60 series, nominal accuracy: ±1.5 mm) was used to obtain the actual distance between people and live areas and live parts. To eliminate operational errors, three independent measurements were performed at each test location, and the average value was taken as the final true value. The experiment preset a graded warning strategy, setting three thresholds of 1 m, 2.5 m, and 4 m, corresponding to "danger" (red), "alarm" (orange), and "caution" (yellow) signals, respectively. When the monitoring distance is lower than the corresponding threshold, the system will trigger the corresponding color signal and simultaneously change the color of the ranging line and the target detection box, intuitively reflecting the risk level.

[0070] like Figure 14 As shown, the red rectangles mark the charged areas, and the red circles mark the specific charged parts. Sub-figures (a)-(d) show the actual detection results of this method. As the staff moves from near to far, the system can accurately locate the key points of the human foot, predict the distance between the person and the charged area and part in real time, realize multi-target detection, and present the results in a visual manner according to the hierarchical early warning strategy.

[0071] The ranging results at different distances are shown in Table 3 below, including the distances predicted by the baseline model between the person and the charged area and location. d pbc-region and d pbc-part .

[0072] Table 3: Field measurement results;

[0073] Table 3 shows that the prediction results of the baseline model ( d pbcIn this scenario, due to the lack of targeted feature enhancement, keypoint detection exhibits significant bias. The mean absolute error (MAE) is 0.201 m in the charged region and 0.159 m in the local area. This is particularly evident in long-distance scenarios. Z c >7 m), due to the blurred features of small targets, the ranging error of the baseline model increases sharply, making it difficult to meet the safety requirements of high-precision operations in substations. After introducing the innovative module in this paper ( d pb Thanks to more accurate keypoint localization capabilities, ranging errors decreased (MAE for regions and locations decreased to 0.162 m and 0.103 m, respectively). This verifies that improved feature extraction capabilities can indeed alleviate some of the ranging bias caused by detection jitter, but simple geometric backprojection still has inherent systematic distortions at long distances. After introducing the LEC model ( d pa The ranging accuracy has achieved a qualitative leap. The average error across the entire field has been reduced to 0.021 m (region) and 0.024 m (location), with standard deviations (STDs) as low as 0.017 m and 0.020 m, respectively. This demonstrates that, through the combined effects of feature-level optimization and LEC geometric correction, the proposed method achieves significant error convergence and improved prediction stability in both near and far fields, validating its effectiveness and practicality in complex substation environments.

Claims

1. A target spatial ranging and operational safety early warning method based on single-view geometry, characterized in that: Includes the following steps: S1. Acquire images of operations in multiple scenarios and supplement them with COCO dataset samples to improve diversity; perform foot key point annotation and data cleaning on all images to construct a dataset of foot key points for substation workers. S2, construct the YOLO-P3C keypoint detector. The detector integrates a multi-scale pyramid structure of feature extraction module CIDW, feature enhancement module CDA, and fusion CEM module to detect foot keypoints; S3, based on the transformation relationship between the world coordinate system, camera coordinate system, image coordinate system and pixel coordinate system, uses the inverse coordinate transformation formula and combines the prior height value to map the two-dimensional image coordinates to the three-dimensional world coordinates and derive the depth information. S4 proposes an error compensation model based on reverse coordinate transformation, LEC, which uses the camera coordinate system depth and the initial world coordinates as geometric priors to construct an explicit compensation function to correct the nonlinear depth deviation in the monocular reconstruction process and obtain the corrected three-dimensional coordinates. S5, based on the corrected three-dimensional coordinates, uses a single-view metrology method to solve the key points of the head, calculates the safe distance between maintenance personnel and live areas and parts, and realizes hierarchical early warning and visualization display according to preset thresholds.

2. The target spatial ranging and operation safety early warning method based on single-view geometry according to claim 1, characterized in that, In step S1, six key points are selected in the dataset: right big toe, right little toe, right heel, left big toe, left little toe, and left heel. Low-resolution, motion-blurred, and severely occluded samples are removed during data cleaning.

3. The target spatial ranging and operation safety early warning method based on single-view geometry according to claim 1, characterized in that, In step S2, the feature extraction module CIDW is implemented as follows: The feature extraction module CIDW is constructed, which draws on multi-branch receptive field design, partial convolution strategy and multi-dimensional attention mechanism. The conventional two-dimensional convolution in the C3k structure is replaced with a partial convolution module, which processes only one-quarter of the input channels. The multi-dimensional attention mechanism MCA is introduced to apply attention in the three dimensions of height, width and channels respectively, which enhances the ability to capture fine-grained features of small targets.

4. The target spatial ranging and operation safety early warning method based on single-view geometry according to claim 1, characterized in that, In step S2, the feature enhancement module CDA is implemented as follows: A feature enhancement module CDA is constructed, which introduces the idea of ​​deformable sampling into the self-attention mechanism and combines the CSP structure with the feedforward network to achieve feature fusion. A lightweight convolutional network is used to predict the offset corresponding to each query point and dynamically adjust the sampling position so that the attention distribution can adaptively focus on the dense foot key point region according to the content.

5. The target spatial ranging and operation safety early warning method based on single-view geometry according to claim 1, characterized in that, In step S2, the integration of the CEM module into the multi-scale pyramid structure is specifically implemented as follows: The CEM module is integrated into a multi-scale pyramid structure. High-resolution semantic feature representation is enhanced by channel-space joint attention and multi-receptive field convolution aggregation. The statistical properties of the input features are extracted using global pooling to generate channel attention weights, spatial statistical properties are calculated to generate spatial attention weights, and depth-separable convolutions of different sizes are used to expand channel features and fuse information from different receptive fields.

6. The target spatial ranging and operation safety early warning method based on single-view geometry according to claim 1, characterized in that, In step S3, 3D reconstruction is achieved by using constant parameters of the camera at a fixed angle, combined with prior height values ​​and the inverse coordinate transformation formula; the specific transformation is as follows: Establish the transformation formula between world coordinates and pixel coordinates: ； in, For depth information; u , v () represents the pixel coordinates of a point on the image plane; X w , Y w , Z w () represents the coordinates of a point in the world coordinate system; R Let be a rotation matrix. T It is a translation vector; Based on the inverse coordinate transformation relationship, the known target height value is introduced. The depth information of the target is derived using the following formula. : ； ； in, , P (2, 0) represents the value in the third row and first column of the matrix; based on the calculated... It can derive three-dimensional world coordinates from two-dimensional pixel coordinates and use the derived three-dimensional world coordinates as geometric prior inputs for subsequent LEC models.

7. The target spatial ranging and operation safety early warning method based on single-view geometry according to claim 1, characterized in that, In step S4, the error compensation model LEC based on reverse coordinate transformation is as follows: Depth of camera coordinate system with initial world coordinates As a geometric prior, an explicit compensation function for long-range errors is constructed; Multinomial regression is used to jointly compensate the error term, with the depth in the camera coordinate system as the basis. and horizontal coordinates As a joint input, a polynomial basis function is constructed to capture the nonlinear coupling of errors in the depth and lateral directions: ； ； in, X e and Y e These are the error compensation values ​​in the X-axis and Y-axis directions, respectively, with coefficients { a i }and{ b i The interaction term was obtained by least squares fitting. Z c X w and quadratic terms Z c 2 , X w 2 Used to capture the nonlinear coupling between lateral coordinates and depth variations; higher-order terms Z c 3 Used to model the trend of depth error increasing with the square or even cube of distance.

8. The target spatial ranging and operation safety early warning method based on single-view geometry according to claim 1, characterized in that, In step S5, the specific method for implementing safe distance monitoring and early warning is as follows: Obtain the target's 3D coordinates: For charged areas, the LEC model is used to correct coordinates; for maintenance personnel, the head coordinates are solved by combining foot key point coordinate mapping and single-view metrology methods, and the mapping relationship between pixel coordinates and physical height is established using the following formula: ； in, η Indicates the scaling factor. p b and p t These represent the coordinates of the operations and maintenance personnel at the bottom and top of the pixel coordinate system, respectively. θ and μ These represent the horizontal cancellation line and the vertical cancellation point in the pixel coordinate system, respectively. Calculate distance and issue a warning: Based on the known world coordinates of the charged parts and the solved personnel coordinates, the safe distance is calculated using the following formula. d : ； Based on the preset graded early warning strategy, different distance thresholds are set. When the monitoring distance is lower than the corresponding threshold, the corresponding color signal and visual warning are triggered.

9. A computer device, characterized in that, It includes a memory and a processor, which are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the target space ranging and operation safety early warning method based on single-view geometry as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the target spatial ranging and operation safety early warning method based on single-view geometry as described in any one of claims 1 to 8.

Images