Pose estimation system and method for distribution network hot-line work robot

Abstract

The invention discloses a distribution network hot-line work robot pose estimation system and method, and belongs to the technical field of robot visual perception. The system comprises an input preprocessing module which is used for carrying out noise reduction and enhancement processing on an RGB-D image; the shared feature extraction module is used for extracting multi-scale universal features based on a lightweight convolution architecture; the 6D pose estimation module is used for processing the image based on the neural implicit field to obtain a pose estimation result; and the joint optimization module is used for realizing detection and pose estimation shared feature extraction through cooperation of a multi-task loss function and a pose estimation result. The system adopts adaptive median filtering and homomorphic filtering to eliminate noise and uneven illumination, and bilateral filtering optimizes a depth map; a bottleneck structure and cavity convolution are introduced into feature extraction, and a rank enhancement linear attention module is embedded; according to the pose estimation, a geometric field and an appearance field are modeled through a neural implicit field, and pose hypotheses are generated and optimized. According to the method, the problems of low pose estimation precision and poor real-time performance in the distribution network live working environment are solved, and the working safety and efficiency of the robot are improved.

Description

Technical Field

[0001] This application relates to the field of robot vision perception technology, specifically to a pose estimation system and method for a live-line working robot in power distribution networks. Background Technology

[0002] With the intelligent development of power systems, live-line working robots in power distribution networks are playing an increasingly important role in the maintenance, repair, and fault handling of power equipment. These robots need to accurately perceive and operate power equipment, which requires them to possess accurate target detection and pose estimation capabilities. Target detection and pose estimation, as core technologies of robot visual perception, directly affect the safety and reliability of the robot in performing live-line working tasks.

[0003] In live-line working scenarios in power distribution networks, target detection technology is mainly used to identify the type, location, and status of power equipment, providing robots with basic information about the targets they operate on. Traditional target detection methods are primarily based on hand-designed feature extractors and classifiers, such as HOG features and SVM classifiers. With the development of deep learning technology, target detection algorithms based on convolutional neural networks, such as the YOLO series and Faster R-CNN, have achieved significant improvements in detection accuracy and speed. However, when these general-purpose target detection algorithms are applied to live-line working scenarios in power distribution networks, they still face problems such as poor adaptability to complex environments, difficulty in detecting small targets, and weak dynamic scene processing capabilities.

[0004] Pose estimation technology is used to determine the position and orientation of power equipment in three-dimensional space, providing spatial reference for precise robot operation. Traditional pose estimation methods are mainly based on feature matching and geometric constraints, such as the PnP algorithm. In recent years, 6D pose estimation methods based on deep learning have gradually become a research hotspot. CN114912287B discloses a robot autonomous grasping simulation system and method based on target 6D pose estimation. This system uses ROS-Gazebo to build a robot vision-assisted grasping simulation system. For target pose estimation, deep learning is used to extract features from RGB-D data and fuse pixel-by-pixel features, and then iteratively refines the data to obtain accurate pose information. However, this method is mainly designed for simulation environments and has limited adaptability in the actual complex live-line working environment of power distribution networks.

[0005] To address the pose estimation problem in complex environments, CN111899301A proposes a deep learning-based 6D pose estimation method for workpieces. This method segments the target object using a semantic segmentation model, fuses 3D point cloud data and RGB information to construct a dense fusion network to estimate the object's 3D position and orientation information, and employs the ICP algorithm for iterative matching and pose fine-tuning. This method solves the 6D pose estimation problem for target objects in complex environments to a certain extent, but improvements are still needed in terms of real-time performance and computational efficiency.

[0006] While the aforementioned methods have made some progress in target detection and pose estimation, live-line work in power distribution is a crucial means of ensuring the safe and stable operation of power systems. However, it faces challenges such as strong electromagnetic interference, complex lighting conditions, and diverse targets (e.g., insulators, clamps, conductors). Traditional methods have the following shortcomings: 1. Insufficient target detection accuracy. Poor adaptability to complex environments: The surface textures of power equipment are similar (such as porcelain insulators and metal parts), and traditional convolutional neural networks are prone to false detections; it is difficult to detect small targets, as key components such as bolts and pins are small in size and are easily ignored in long-distance images; it has weak dynamic scene processing capabilities: when wind causes the conductor to sway, traditional methods are difficult to track and locate in real time.

[0007] 2.6D pose estimation is inaccurate. Lacking geometric constraints, existing methods rely heavily on 2D image features, making it difficult to accurately recover the target's 3D pose (position + rotation); low illumination robustness, with unstable feature extraction under backlighting or strong light reflection conditions; insufficient fusion of semantic and geometric information, resulting in ineffective collaboration between detection results and pose estimation, leading to accumulated localization errors. 3. The conflict between real-time performance and robustness. Power operations have high safety requirements, necessitating both high accuracy and low latency, but existing algorithms struggle to achieve this balance.

[0008] Therefore, there is an urgent need to develop a pose estimation system for live-line working scenarios in power distribution networks. Through feature sharing and multi-task collaborative optimization, the system's detection accuracy, pose estimation accuracy, and real-time performance in complex environments can be improved, providing reliable visual perception capabilities for live-line working robots in power distribution networks. Summary of the Invention

[0009] To address the technical problems of insufficient target detection accuracy, inaccurate 6D pose estimation, and the contradiction between real-time performance and robustness in live-line working environments for power distribution, this application provides a pose estimation system and method for live-line working robots in power distribution networks.

[0010] The technical solution adopted by this application to solve its technical problem is: In a first aspect, this application provides a pose estimation system for a live-line working robot in a power distribution network, comprising: The input preprocessing module is used to perform noise reduction and enhancement processing on the RGB-D images acquired by the live-line working robot of the power distribution network to obtain preprocessed images; A shared feature extraction module is used to extract general features containing multi-scale information in preprocessed images based on a lightweight convolutional architecture. The 6D pose estimation module is used for general feature processing of preprocessed images containing multi-scale information based on neural implicit fields to obtain 6D pose estimation results.

[0011] Furthermore, it also includes a joint optimization module, which is used to coordinate the 6D pose estimation results through a multi-task loss function, detect parameters of the feature extraction module that are shared with the pose estimation, and synchronously optimize the 6D pose estimation module through backpropagation.

[0012] Furthermore, the input preprocessing module includes: The image denoising unit is used to suppress salt-and-pepper noise in power distribution scenarios by using adaptive median filtering for RGB-D images, and to eliminate the influence of uneven illumination by combining homomorphic filtering to obtain the denoised image; The depth map optimization unit is used to smooth the depth values ​​of the denoised image through bilateral filtering, fill in invalid pixels, and obtain a preprocessed image.

[0013] Furthermore, the shared feature extraction module is specifically used for: Multi-stage feature extraction: shallow features extract basic information; mid-level features introduce a bottleneck structure (Bottleneck) and combine dilated convolution to capture medium-sized targets; deep features employ a cross-stage local connectivity structure to output high-level features. Feature enhancement: Embedding RALA modules in the backbone network.

[0014] Furthermore, the shared feature extraction module uses YOLOv12 as the core architecture for object detection. In YOLOv12, RALA is embedded between the convolutional layers in the Backbone part, residual connections are introduced in the downsampling module C3k2, RALA is used in the Neck part, and a decoupled design is adopted in the Head part, along with a specific loss function. Specifically, in the Backbone part, the input preprocessed image first goes through multiple convolutional layers for preliminary feature extraction; in the original YOLOv12 Backbone, a rank-enhanced linear attention module is embedded after some convolutional layers. The Neck section fuses feature maps of different scales. Specifically, it uses the RALA attention mechanism in the A2C2f module. For the feature maps in the A2C2f module, RALA is applied to each block region according to a block-based strategy, with a computational complexity of O(n log n). , where N is the total number of pixels in the feature map; In terms of upsampling and stitching, the Upsample module is used to upsample low-resolution feature maps to high-resolution ones. It adopts an adaptive upsampling method, which dynamically adjusts the upsampling weights according to the content of the feature map and the distribution of the target. The head section employs a decoupled design, consisting of a classification branch and a regression branch; the classification branch uses FocalLoss, calculated as follows: ,in As a category balance factor, To focus on the parameters, the regression branch uses DIoU loss combined with CIoU improvement terms. ,in .

[0015] Furthermore, the rank-enhanced linear attention module is used for: The rank enhancement mechanism in the KV buffer fuses detection semantics and spatial features to generate context-aware global queries. :

[0016] Where c is the detection category label, which is mapped to a 64-dimensional vector through a learnable embedding layer; then the weight modulation is calculated. , ,in It is a nonlinear kernel function; in the rank recovery of the output features, a channel interaction matrix is ​​introduced. Features of the original token Perform linear transformation , ; The downsampling module C3k2 is used to perform downsampling operations. In the C3k2 module, the input features are divided into two paths. One path goes through a series of convolution operations, and the other path is directly connected to the output of the module. Finally, the results of the two paths are added together.

[0017] Furthermore, the target detection branch includes: The multi-scale feature fusion unit is used to upsample and concatenate the features output from different levels of the backbone network through the neck network to generate a fusion feature map containing rich details and semantics. An anchor box mechanism is applied to the feature map to predict the class probability of the target and the coordinates of the 2D detection box. The loss function unit uses CIoU loss to calculate the regression error of the detection box.

[0018] Furthermore, the 6D pose estimation module is used for: 1) Neural implicit field modeling, including: geometric field, which takes feature vectors and 3D spatial point coordinates as input and outputs signed distance field values ​​to describe the geometry of the target surface; appearance field, which combines 3D point normal information to predict the RGB color of the corresponding point and extract illumination invariant features. 2) Attitude hypothesis generation and optimization, including: Initial assumptions are made based on prior knowledge of the detection box center and target size, generating an initial pose set that includes random rotations and translations; Hypothesis filtering: A virtual view is synthesized through neural implicit field rendering, and its feature matching degree is calculated with the input image to filter out low-confidence hypotheses; Iterative optimization is performed using the gradient descent algorithm, with the SDF value and image reprojection error as the optimization targets, to iteratively adjust the pose parameters.

[0019] Furthermore, the 6D pose estimation module incorporates the FoundationPose network structure. In the data generation part, synthetic data is generated using language and diffusion models in conjunction with detection results, and the rendering region and physical simulation parameters are optimized. In the neural object modeling part, reference images are preprocessed using detection bounding boxes, the network structure is fine-tuned, and RALA is introduced. Multi-task training is performed using detection auxiliary information. In the pose hypothesis generation part, the sampling strategy is optimized based on detection semantic information, RALA is used to enhance feature expression, and pose updates are improved by combining detection scale information. In the pose selection part, RALA is introduced into the encoder, and the pose ranking is adjusted in conjunction with detection confidence, and multiple hypothesis fusion and screening are performed.

[0020] Furthermore, the 6D pose estimation module specifically includes: In FoundationPose, based on the target category information output by the detection module, an LLM is used to generate a description of the target's appearance. This description is then input into the diffusion model to enhance the appearance of the original model, generating models with diverse appearances. From the perspective of the physics engine and path tracing, combined with the detected target location information, the rendering area is precisely located and cropped. Only the region of interest containing the target is rendered. Based on the target's geometric features and material properties, the simulation parameters in the physics engine are optimized to make the rendered image more realistic. Let the target's geometric feature parameter be G, the material property parameter be M, and the rendering area parameter be R. The physics engine parameters are adjusted using the function f(G,M,R). In the model-free setting of Neural Object Modeling, based on the FoundationPose architecture, a neural RGBD object field is constructed using a small number of reference images. The reference images are preprocessed using the target bounding box information provided by the detection module, and the target regions in the reference images are enlarged and enhanced. The resulting image features are then input into the geometry network and the appearance network. The structure of the geometry network and the appearance network is fine-tuned, and a RALA module is introduced into the network. In a geometric network, let the input features be... The output after RALA processing Through a series of convolution operations and RALA operation The calculation process is as follows: This enhances the network's ability to extract the geometric features of the target. In the appearance network, let the input features be... The output is ,go through and Operation, i.e. This improves the ability to extract target appearance features, thereby enhancing the modeling accuracy of the neural implicit field. When training the geometry and appearance networks, a multi-task learning approach is adopted, incorporating auxiliary information from the target detection task as additional supervision signals into the network training process. Let the auxiliary information for the detection task be... The geometric network loss function is The appearance network loss function is The total loss function is ,in These are the weighting coefficients; In the pose hypothesis generation stage, the global pose sampling method based on FoundationPose is optimized by combining the detected target semantic information to generate pose hypotheses. For symmetrical targets, reduce repeated sampling in the symmetrical direction; for targets with specific functional structures, increase the sampling density in their functionally relevant directions; in the encoder section, the RALA module is used to process the input features; let the input features be... The output after RALA processing is ,pass Enhanced features: When updating rotation and translation via the Transformer encoder, the target scale information S output by the detection module is introduced to adjust the pose update parameters; let the rotation parameter be r and the translation parameter be t, then r and t are adjusted using the function g(S). , During the update process, a hierarchical contrastive network method is used to compare and filter pose assumptions at different levels, prioritizing the retention of assumptions that are closer to the actual pose of the target. In the Pose Ranking Encoder, the structure of FoundationPose is borrowed, and a RALA module is introduced to enhance the extraction and comparison capabilities of features from different pose assumptions. RALA is applied in the convolutional residual blocks and self-attention layers of the encoder. Let the input pose assumption features be... The output after RALA processing is ,pass This allows the model to better capture the differences between different pose assumptions; when calculating the pose ranking, the target confidence information C provided by the detection module is used to weight and adjust the ranking of the pose assumptions; let the initial ranking score of the pose assumption be... The weighted score is Adjusted by the function h(C), i.e. ; When processing multiple pose hypotheses, a multi-hypothesis fusion strategy is adopted to fuse the features of different pose hypotheses, and then feature interaction and filtering are performed through a self-attention mechanism. A constraint based on the target detection results is also introduced for pose hypotheses that are obviously inconsistent with the detected target position and appearance.

[0021] Secondly, this application provides a pose estimation method for a live-line working robot in a power distribution network, based on the pose estimation system for the live-line working robot in a power distribution network, characterized in that it includes: The RGB-D images acquired by the live-line working robot in the power distribution network are denoised and enhanced to obtain pre-processed images; Based on a lightweight convolutional architecture, general features containing multi-scale information in preprocessed images are extracted. Based on the general feature processing of preprocessed images containing multi-scale information using neural implicit fields, 6D pose estimation results are obtained.

[0022] It also includes: coordinating 6D pose estimation results through a multi-task loss function, detecting shared feature extraction module parameters between pose estimation and detection, and synchronously optimizing 6D pose estimation results through backpropagation.

[0023] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the pose estimation method for a live-line working robot of claim 10.

[0024] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the pose estimation method for a live-line working robot in claim 10.

[0025] Fifthly, this application provides a computer program product, the computer program product including computer instructions, the computer instructions instructing a computer to execute the pose estimation method for a live-line working robot in claim 10.

[0026] The beneficial effects of this application are as follows: The pose estimation system for live-line working robots in this application mainly includes: a shared feature extraction module, which extracts multi-scale features of the target based on an improved YOLO architecture; a rank-enhanced attention module, which enhances feature representation capabilities through tensor decomposition technology, paying particular attention to key structural features of power equipment (such as conductor edges and insulator string connection points); and a 6D pose estimation module, which incorporates improvements to the FoundationPose network structure and a joint optimization module, designing a multi-task loss function to collaboratively optimize detection and pose estimation tasks. This system achieves problem-oriented fusion, addressing the issues of attention, pose estimation, and multi-task fragmentation in traditional visual perception. It organically integrates YOLOv12, FoundationPose, and RALA methods into a unified framework. The performance improvement driven by rank-enhanced attention involves applying RALA to relevant modules in YOLOv12 and FoundationPose, enhancing the feature representation capabilities of detection and pose estimation by increasing feature rank, thus overcoming the performance bottleneck of traditional methods. Deep collaboration between detection and pose estimation: From data generation and feature extraction to model training, the entire process of detection and pose estimation is collaboratively implemented. By sharing the feature backbone and joint loss function, the efficiency of feature reuse across tasks and the overall performance are improved. Attached Figure Description

[0027] Figure 1 A schematic diagram of a pose estimation system for a live-line working robot in a power distribution network, provided in this application; Figure 2 A schematic diagram of an electronic device provided in this application. Detailed Implementation

[0028] The technical solution of this application will now be clearly and completely described with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0029] In live-line working scenarios in power distribution networks, live-line working robots face core challenges such as difficulty in visual perception of complex power equipment, insufficient accuracy of 6D pose estimation, and poor coordination between detection and pose tasks. Traditional methods, due to limitations in attention mechanisms, weak feature representation capabilities, and fragmentation of multi-task operations, are unable to meet the stringent requirements of "accurate perception, reliable positioning, and efficient coordination" in live-line working. To address these pain points, this application proposes the following solution: This application relates to the interdisciplinary field of computer vision, robotics, and artificial intelligence, specifically to real-time target detection, 6D pose estimation, and multi-task joint optimization techniques for live-line working robots in power distribution networks. By fusing Rank-Enhanced Linear Attention (RALA), Neural Implicit Field, and hierarchical contrastive networks, combined with a lightweight YOLOv12 network structure, an end-to-end unified framework is constructed to resolve the contradiction between target perception accuracy and real-time performance in scenarios such as strong electromagnetic interference, complex lighting, and multi-scale changes in targets during live-line working in power distribution networks. It is particularly suitable for tasks such as autonomous operation and inspection of live-line working robots, power equipment defect identification, and operation path planning, while also accommodating general vision applications such as intelligent inspection, industrial robotic arm grasping, and AR scene reconstruction.

[0030] Among them, the YOLO (You Only Look Once) series of object detection algorithms is known for its real-time performance and efficiency. Its core design philosophy is single-stage detection (without region proposal) and end-to-end training. YOLOv12 includes a backbone network and a head.

[0031] Rank-enhanced linear attention (RALA) improves model expressiveness while maintaining linear complexity by explicitly introducing rank-constrained optimization or dynamic rank expansion.

[0032] A neural implicit field is a type of field that uses neural networks (such as multilayer perceptrons, MLPs) to represent spatial coordinates (such as 3D points). x , y , z Implicit fields are representations that map spatial properties to attribute values ​​(such as color, density, and semantic labels). Unlike traditional explicit representations (such as voxel meshes, point clouds, and meshes), implicit fields directly model spatial properties through parameterized functions, without the need to explicitly store geometric or texture information.

[0033] Hierarchical Contrastive Network is a deep learning framework that combines multi-scale feature extraction with contrastive learning. It aims to improve the model's semantic understanding and generalization ability of data by constructing hierarchical contrastive tasks.

[0034] This application proposes a pose estimation system and method for live-line working robots in power distribution networks, which enables high-precision, real-time target detection and 6D pose estimation in live-line working scenarios, thereby improving the robot's autonomous operation capabilities.

[0035] Example 1 like Figure 1 As shown, a pose estimation system for a live-line working robot in a power distribution network includes an input preprocessing module, a shared feature extraction module, a 6D pose estimation module, and a joint optimization module.

[0036] The input preprocessing module is used to perform noise reduction and enhancement processing on the RGB-D images acquired by the live-line working robot of the power distribution network to obtain preprocessed images.

[0037] Among them, RGB-D images are an important data format in the field of computer vision, formed by fusing RGB color images and depth images. An RGB-D image consists of two core parts: the RGB color image, which presents color information through the superposition of red (R), green (G), and blue (B) channels, covering most colors perceptible to human vision; and the depth map, stored as a grayscale image, where each pixel value represents the actual distance from the sensor to the object (in millimeters). The depth map is strictly registered with the RGB image, with each pixel corresponding to the next.

[0038] The input preprocessing module includes an image denoising unit and a depth map optimization unit. The image denoising unit, for RGB-D images, employs adaptive median filtering to suppress salt-and-pepper noise in power distribution scenarios, and combines this with homomorphic filtering to eliminate the effects of uneven illumination, resulting in a denoised image.

[0039] Adaptive median filtering dynamically adjusts the filter window size based on noise density, using a larger window for areas with high noise density and a smaller window for areas with low noise density, effectively preserving image details while removing salt-and-pepper noise. Homomorphic filtering performs frequency domain transformation on the image, separating the illumination and reflection components, compressing the illumination component and enhancing the reflection component, thereby eliminating shadows and highlights caused by uneven lighting in power distribution scenes. The depth map optimization unit smooths the depth values ​​of the denoised image using bilateral filtering, filling in invalid pixels to obtain a preprocessed image. Bilateral filtering considers both spatial distance and pixel value differences, smoothing the depth map while preserving edge information and avoiding blurring of target boundaries. For invalid pixel areas in the depth map, an interpolation algorithm based on surrounding valid pixels is used to fill in the gaps, ensuring the integrity of the depth information.

[0040] The shared feature extraction module is used to extract general features containing multi-scale information from preprocessed images based on a lightweight convolutional architecture. The shared feature extraction module performs multi-stage feature extraction: (a) shallow features, extracting basic information; (b) mid-level features, introducing a bottleneck structure (Bottleneck) and combining dilated convolutions to capture medium-sized targets; (c) deep features, employing a cross-stage local connectivity structure to output high-level features; feature enhancement is embedded in the RALA module of the backbone network. Shallow feature extraction is mainly achieved through standard convolutional layers with a kernel size of 3×3 and a stride of 1, extracting basic visual features such as edges and textures of the image. Mid-level feature extraction introduces the Bottleneck structure, using 1×1 convolutions for dimensionality reduction, 3×3 convolutions for feature extraction, and then 1×1 convolutions for dimensionality enhancement, reducing computation while maintaining feature expressiveness. Dilated convolutions expand the receptive field by inserting holes in the convolution kernel, effectively capturing features of medium-sized targets; the dilation rate is set to 2. Deep feature extraction employs a cross-stage local connectivity structure to fuse shallow and deep features, thereby enhancing the expressive power of the features.

[0041] The shared feature extraction module uses YOLOv12 as the core architecture for object detection. In YOLOv12, RALA is embedded between the convolutional layers in the Backbone part, residual connections are introduced by the downsampling module C3k2, RALA is used in the Neck part, and a decoupled design is adopted in the Head part, with a specific loss function designed. Specifically: (1) In the Backbone part, the input preprocessed image first goes through multiple convolutional layers for preliminary feature extraction; in the Backbone of the original YOLOv12, a rank-enhanced linear attention module is embedded after some convolutional layers. The rank-enhanced linear attention module uses the rank enhancement mechanism in the KV buffer to fuse detection semantics and spatial features to generate a context-aware global query. , where c is the detection category label, mapped to a 64-dimensional vector through a learnable embedding layer. Then, weight modulation is calculated. , ,in Using a non-linear kernel function to ensure non-negativity of weights, this design improves the rank of the KV buffer from 58% to 92% compared to traditional linear attention, and increases feature diversity by 40%. Secondly, it restores the rank of the output features by introducing a channel interaction matrix. Features of the original token Perform linear transformation , By enhancing inter-channel correlation through the Hadamard product, the output feature rank is further increased from 92% to 98% (the full rank of a matrix is ​​defined as...). (Approaching the Softmax attention level); the downsampling module C3k2 is used to perform downsampling operations. In the C3k2 module, the input features are divided into two paths. One path goes through a series of convolution operations, and the other path is directly connected to the output of the module. Finally, the results of the two paths are added together.

[0042] The Neck section fuses feature maps of different scales. Specifically, it uses the RALA attention mechanism in the A2C2f module. For the feature maps in the A2C2f module, RALA is applied to each block region according to a block-based strategy, with a computational complexity of O(n log n). Where N is the total number of pixels in the feature map. Regarding upsampling and stitching, the Upsample module is used to upsample low-resolution feature maps to high-resolution ones. It employs an adaptive upsampling method, dynamically adjusting the upsampling weights based on the content of the feature map and the distribution of the target.

[0043] The head section employs a decoupled design, divided into classification and regression branches. The classification branch uses FocalLoss, calculated as follows: ,in As a category balance factor, To focus on parameters and address class imbalance, thus improving detection accuracy for minority class targets, the regression branch employs DIoU loss combined with a CIoU improvement term. ,in This improves the accuracy of bounding box regression, especially the accuracy of aspect ratio regression.

[0044] The object detection branch includes a multi-scale feature fusion unit and a loss function unit. The multi-scale feature fusion unit upsamples and concatenates features from different levels of the backbone network output through the neck network, generating a fused feature map rich in detail and semantics. Anchor boxes are then applied to this feature map to predict the object's class probability and 2D bounding box coordinates. The loss function unit calculates the bounding box regression error using CIoU loss.

[0045] The 6D pose estimation module is used for general feature processing of preprocessed images containing multi-scale information based on neural implicit fields to obtain 6D pose estimation results. The 6D pose estimation module performs neural implicit field modeling, including: (a) geometric field, which takes input feature vectors and 3D spatial point coordinates, and outputs signed distance field (SDF) values ​​to describe the geometry of the target surface; (b) appearance field, which combines 3D point normal information to predict the RGB color of corresponding points and extract illumination-invariant features. The 6D pose estimation module also performs pose hypothesis generation and optimization, including: (a) initial hypothesis, which generates an initial pose set containing random rotations and translations based on prior knowledge of the detection box center and target size; (b) hypothesis filtering, which synthesizes a virtual view through neural implicit field rendering, calculates feature matching degree with the input image, and filters low-confidence hypotheses; and (c) iterative optimization, which uses the gradient descent algorithm to iteratively adjust pose parameters with SDF values ​​and image reprojection error as optimization targets.

[0046] The 6D pose estimation module incorporates the FoundationPose network structure. In the data generation part, synthetic data is generated using language and diffusion models in conjunction with detection results, and the rendering region and physical simulation parameters are optimized. In the neural object modeling part, reference images are preprocessed using detection bounding boxes, the network structure is fine-tuned, and RALA is introduced. Multi-task training is performed using detection auxiliary information. In the pose hypothesis generation part, the sampling strategy is optimized based on detection semantic information, RALA is used to enhance feature representation, and the pose update is improved by combining detection scale information. In the pose selection part, RALA is introduced into the encoder, and the pose ranking is adjusted in conjunction with detection confidence to perform multi-hypothesis fusion and screening.

[0047] The 6D pose estimation module in FoundationPose uses LLM to generate a description of the target's appearance based on the target category information output by the detection module. This description is then input into the diffusion model to enhance the appearance of the original model, generating models with diverse appearances. From the perspective of the physics engine and path tracing, combined with the detected target position information, the rendering area is precisely located and cropped. Only the region of interest containing the target is rendered. Based on the target's geometric features and material properties, the simulation parameters in the physics engine are optimized to make the rendered image more realistic. Let the target's geometric feature parameter be G, the material property parameter be M, and the rendering area parameter be R; the physics engine parameters are adjusted using the function f(G,M,R).

[0048] In the model-free setting of Neural Object Modeling, a neural RGBD object field is constructed using a small number of reference images based on the FoundationPose architecture. The reference images are preprocessed using target bounding box information provided by the detection module, which enlarges and enhances the target regions within the reference images. The resulting image features are then input into the geometric network. and appearance network Fine-tune the structure of the geometric and appearance networks, and introduce RALA modules into the networks.

[0049] In a geometric network, let the input features be... The output after RALA processing Through a series of convolution operations and RALA operation The calculation process is as follows: This enhances the network's ability to extract the geometric features of the target. Similarly, in the appearance network, let the input features be... The output is ,go through and Operation, i.e. This improves the ability to extract target appearance features, thereby enhancing the modeling accuracy of the neural implicit field. When training the geometry and appearance networks, a multi-task learning approach is adopted, incorporating auxiliary information from the target detection task (such as the approximate shape and color distribution of the target) as additional supervision signals into the network training process. Let the auxiliary information for the detection task be... The geometric network loss function is The appearance network loss function is The total loss function is ,in These are weight coefficients, which enable the network to better coordinate with the detection task when learning the geometric and appearance features of the target, thereby improving the network's generalization ability and adaptability to new targets. In the pose hypothesis generation stage, the global pose sampling method based on FoundationPose, combined with detected target semantic information, optimizes the generation of pose hypotheses. Different pose sampling strategies are adopted for different target categories based on the detected target category: for symmetrical targets, repeated sampling in the symmetrical direction is reduced; for targets with specific functional structures, the sampling density in their functionally relevant directions is increased. In the encoder part, the RALA module is used to process the input features; let the input features... The output after RALA processing is ,pass Enhancing the expressive power of features makes the generated pose hypotheses more diverse and accurate. When updating rotation and translation using the Transformer encoder, the target scale information S output by the detection module is introduced to adjust the pose update parameters. Let the rotation parameter be r and the translation parameter be t; the function g(S) is used to adjust r and t. , This makes attitude updates more precise and accurate.

[0050] During the update process, a hierarchical contrastive network method is used to compare and filter pose assumptions at different levels, prioritizing those that are closer to the actual target pose. In the Pose Ranking Encoder, the structure of FoundationPose is borrowed, and a RALA module is introduced to enhance the extraction and comparison capabilities of features from different pose assumptions. RALA is applied in the convolutional residual blocks and self-attention layers of the encoder; let the input pose assumption features be... The output after RALA processing is ,pass This allows the model to better capture the differences between different pose assumptions; when calculating the pose ranking, the target confidence information C provided by the detection module is used to weight and adjust the ranking of the pose assumptions; let the initial ranking score of the pose assumption be... The weighted score is Adjusted by the function h(C), i.e. .

[0051] When processing multiple pose hypotheses, a multi-hypothesis fusion strategy is adopted to fuse the features of different pose hypotheses, and then feature interaction and filtering are performed through a self-attention mechanism. A constraint based on the target detection results is also introduced to reduce the weight or directly exclude pose hypotheses that are obviously inconsistent with the detected target position and appearance.

[0052] The joint optimization module coordinates the 6D pose estimation results through a multi-task loss function, detects shared feature extraction module parameters, and synchronously optimizes the 6D pose estimation module through backpropagation. The joint optimization module designs a comprehensive loss function comprising two parts: object detection loss and pose estimation loss. The object detection loss includes classification loss and bounding box regression loss, while the pose estimation loss includes rotation loss and translation loss. By adjusting the weights of each loss component, the optimization objectives of different tasks are balanced. During training, the joint optimization module simultaneously updates the parameters of both the shared feature extraction module and the 6D pose estimation module, enabling the two tasks to mutually promote each other and improve overall performance.

[0053] As a specific solution, in this embodiment, the shared feature backbone adopts a variant of CSPDarknet and includes a RALA module in the first 10 layers. The detection and pose estimation branches share the parameters of the first 8 layers. This is based on the similarity between the two tasks in the early stages of feature extraction, realizing feature reuse and reducing redundant computation. The multi-task joint loss function comprehensively considers factors such as detection, pose estimation, feature rank regularization, and weight decay. By reasonably setting the weights, the detection and pose estimation tasks promote each other and optimize together during training, solving the problem of feature fragmentation in multi-task tasks.

[0054] The specific scheme is as follows: (1) Shared feature backbone: The first 10 layers adopt a variant of CSPDarknet, which contains 5 layers of RALA modules, and the output feature map size is . The detection branch and the pose estimation branch share the first 8 layers of parameters, while only the detection head and the pose refinement network are independent. This design is based on the similarity between the two tasks in the early stage of feature extraction, so as to realize feature reuse and reduce redundant calculations; (2) multi-task multi-joint loss function. ,in The detection loss (classification + regression) has a weight of 1. The classification loss uses Focal Loss, and the regression loss uses DIoU + CIoU loss, calculated as described above. The L2 loss is used for pose estimation, with a weight of 3 (more sensitive to pose tasks), and is used to measure the difference between the estimated pose and the true pose. The constraint of feature rank ensures the diversity and expressive power of features. The weights are decayed to a coefficient of 0.001 to prevent overfitting. By appropriately setting the weights of each part, the detection and pose estimation tasks mutually promote each other and optimize together during training, thus solving the problem of feature fragmentation in multi-task tasks.

[0055] Based on the above solutions, this application provides a multi-method fusion framework based on the operational scenario. Traditional visual perception frameworks struggle to meet the cross-task requirements of "equipment detection" and "robotic arm pose estimation" in live-line working robots for power distribution networks, resulting in high false negative rates and excessive pose estimation errors for complex power components such as insulators and clamps. This application integrates YOLOv12 (object detection), FoundationPose (pose estimation), and RALA (rank-enhanced attention) to construct a unified perception framework for live-line working robots for power distribution networks. Through cross-task process reconstruction, multimodal images (visible light + infrared) of power equipment are introduced in the data preprocessing stage, and a shared backbone network is designed in the feature extraction layer to achieve end-to-end optimization from "operation scenario image input" to "joint output of detection boxes and 6D pose".

[0056] Furthermore, in the operation scenarios of live-line working robots in power distribution networks, interference such as surface reflections and occlusions on equipment can easily lead to "low-rank degradation" of feature maps. Traditional attention mechanisms (such as CBAM and SENet) are difficult to effectively capture long-distance dependencies, resulting in detection box offsets and pose angle prediction errors. This application embeds the RALA (Rank-Aware Local Attention) module into the feature encoder of YOLOv12's Neck network and FoundationPose. Through the feature rank enhancement algorithm, low-rank decomposition and rank constraint optimization are performed on the feature maps of key areas of power equipment (such as clamp connections and insulator skirts), enhancing the discriminativeness and robustness of the features.

[0057] Furthermore, this application achieves end-to-end collaboration between detection and pose estimation. In traditional multi-task frameworks, detection and pose branches are trained independently, resulting in low feature utilization. Moreover, the "detection results not calibrated with initial pose values" in the workflow can easily lead to accumulated errors. This application constructs a dedicated dataset for live-line operation robots in power distribution networks, builds labeled "equipment image-robotic arm pose" pairs, and generates extreme lighting and complex occlusion scene data through simulation rendering to enhance the model's generalization ability. Feature extraction is collaborative, and a cross-task feature interaction module is designed. "Detection guidance vectors" and "pose attention masks" are introduced into the output layer of the shared backbone network to achieve feature complementarity between the detection and pose branches. Training and optimization are collaborative, and a multi-task joint loss function is proposed, and a cross-task consistency constraint term is introduced to force the center point of the detection box to align with the coordinates of the pose estimation origin.

[0058] Example 2 like Figure 1 As shown, this application provides a pose estimation system for a live-line working robot in a power distribution network based on rank-enhanced attention, comprising: The input preprocessing module performs noise reduction and enhancement on RGB-D images to adapt to complex working environments (such as electromagnetic interference and backlighting in live power distribution scenarios). The shared feature backbone network, based on a lightweight convolutional architecture, extracts general features containing multi-scale information; specifically, it includes: a rank-enhanced attention module (RALA) and an object detection branch; The Rank-Enhanced Attention Module (RALA) enhances the feature representation of key target structures (such as insulator skirts and conductor textures) through feature tensor decomposition and attention mechanisms. The target detection branch outputs the target category and 2D detection box; the 4.6D pose estimation branch: realizes the regression from 2D features to 3D pose based on the Neural Implicit Field (NIF); Joint optimization module: Collaboratively optimizes the detection and pose estimation tasks through multi-task loss functions.

[0059] The specific implementation process of the pose estimation method for live-line working robots in power distribution networks based on rank-enhanced attention includes: 1. Input preprocessing. (1) Image noise reduction: Adaptive median filtering is used to suppress salt-and-pepper noise in the power distribution scene, and homomorphic filtering is used to eliminate the influence of uneven illumination; (2) Depth map optimization: The depth value is smoothed by bilateral filtering, and invalid pixels (such as distant blurred areas) are filled in to improve the accuracy of 3D geometry calculation; 2. Shared feature backbone network, including: (1) Multi-stage feature extraction: (a) Shallow features. Basic information such as edges and textures are extracted through 3×3 convolution; (b) Mid-level features. A bottleneck structure is introduced to reduce computation, and dilated convolution is combined to expand the receptive field and capture medium-sized targets such as insulator strings and wire clamps; (c) Deep features: Cross-stage local connectivity (CSP) structure is used to enhance gradient flow and output semantically rich high-level features (such as abstract concepts such as "insulator" and "wire"); (2) Feature enhancement: Embed RALA modules in key layers of the backbone network.

[0060] 3. Object detection branch, including: (1) Multi-scale feature fusion. The different levels of features output by the backbone network are upsampled and concatenated through the neck network to generate a fusion feature map containing rich details and semantics. Anchor boxes are applied to the feature map to predict the class probability of the target and the coordinates of the 2D detection box; (2) Loss function: CIoU loss is used to calculate the regression error of the detection box, and Focal Loss is combined to alleviate the problem of imbalance between positive and negative samples; 4.6D pose estimation branch, including: (1) Neural implicit field modeling. (a) Geometric field. Input feature vectors and 3D spatial point coordinates, output signed distance field (SDF) values ​​to describe the geometry of the target surface; (b) Appearance field. Combine 3D point normal information to predict the RGB color of the corresponding point and achieve illumination invariant feature extraction; (2) Pose hypothesis generation and optimization. (a) Initial hypothesis. Based on the prior knowledge of the detection box center and target size, an initial pose set including random rotation and translation is generated; (b) Hypothesis screening: a virtual view is synthesized through neural implicit field rendering, and the feature matching degree is calculated with the input image to filter low-confidence hypotheses; (c) Iterative optimization. Using the gradient descent algorithm, the pose parameters (rotation matrix R, translation vector t) are iteratively adjusted with the SDF value and image reprojection error as the optimization objectives. Step 5. Joint optimization: The detection and pose estimation branches share the backbone network parameters and are optimized synchronously through backpropagation to achieve the synergistic effect of "detection-guided pose and pose-constrained detection".

[0061] Based on the solution in this embodiment, experiments were conducted, and the following experimental results were obtained: 1. Detection performance: Compared with the original YOLOv12-N, the improved RALA-YOLOv12-N improves mAP@0.5:0.95 from 40.6% to 42.1%, and mAP@0.5 from 56.7% to 58.9%, verifying that RALA effectively improves detection accuracy in the YOLOv12 architecture, while maintaining a high frame rate to meet real-time detection requirements; 2. Pose estimation accuracy (YCB-Video, 16 reference images): Compared with FoundationPose, the method in this application reduces the ADD-S error from 85.3mm to a lower level, significantly shortens the rendering time from 1240ms to 520ms, and significantly improves the success frame rate, indicating that by incorporating FoundationPose and combining detection guidance and RALA optimization, the accuracy and efficiency of pose estimation are effectively improved. 3. Collaborative optimization gains: The shared backbone reduces redundant computation by 40%, and the overall FLOPs decrease from 1820G to 1150G. After joint training, the detection mAP is further improved by 0.8%, and the pose ADD-S error is reduced by another 3.1%. This fully demonstrates the collaborative gain effect of detection and pose estimation tasks under the joint optimization framework, and verifies the effectiveness of the fusion of the three methods in this application.

[0062] Example 3 A method for estimating the pose of a live-line working robot in a power distribution network includes the following steps: The RGB-D images acquired by the live-line working robot in the power distribution network are denoised and enhanced to obtain a preprocessed image. This step specifically includes: for the RGB-D image, adaptive median filtering is used to suppress salt-and-pepper noise in the power distribution scene, combined with homomorphic filtering to eliminate the influence of uneven illumination, resulting in a denoised image; the denoised image is then smoothed for depth values ​​using bilateral filtering to fill in invalid pixels, resulting in the preprocessed image. Adaptive median filtering automatically adjusts the filtering window size according to the noise density of local image regions, with a window size ranging from 3×3 to 9×9. Homomorphic filtering performs a logarithmic transformation on the image, converting multiplicative illumination effects into additive effects, then suppresses low-frequency components and enhances high-frequency components in the frequency domain, finally performing an exponential transformation to restore the image. Bilateral filtering considers both spatial distance and pixel value differences, with a spatial kernel standard deviation set to 5 and a value range kernel standard deviation set to 20, effectively preserving depth map edge information while smoothing depth values. For invalid pixels in the depth map, a weighted average interpolation method based on the nearest valid pixel is used for filling.

[0063] Based on a lightweight convolutional architecture, this step extracts general features containing multi-scale information from preprocessed images. Using YOLOv12 as the foundation, it performs multi-stage feature extraction: shallow feature extraction uses standard convolutional layers with a kernel size of 3×3, a stride of 1, and the number of channels increasing progressively from 32; mid-layer feature extraction introduces a Bottleneck structure, using 1×1 convolutions for dimensionality reduction (reduction ratio of 4), 3×3 convolutions for feature extraction, and then 1×1 convolutions for dimensionality increase, combined with dilated convolutions (dilation rate of 2) to expand the receptive field; deep features employ a cross-stage local connectivity structure, fusing shallow and deep features through skip connections. A RALA module is embedded between the convolutional layers of the backbone network to enhance feature representation. The RALA module uses a rank enhancement mechanism in the KV buffer to fuse detection semantic and spatial features, generating a context-aware global query; then, weight modulation is calculated to weight the features; finally, in the rank recovery stage of the output features, a channel interaction matrix is ​​introduced to linearly transform the original token features. The downsampling module C3k2 divides the input features into two paths. One path undergoes a series of convolutional operations (including 1×1 convolution for dimensionality reduction, 3×3 convolution with a stride of 2 for downsampling, and 1×1 convolution for dimensionality increase). The other path performs downsampling through 1×1 convolution and 2×2 max pooling. Finally, the results of the two paths are summed. The Neck part uses the RALA attention mechanism in the A2C2f module, applying RALA to each block of the feature map according to an 8×8 block strategy. The upsampling module employs an adaptive upsampling method, dynamically adjusting the upsampling weights based on the content of the feature map and the distribution of the target. The Head part's detection head adopts a decoupled design, divided into classification and regression branches. The classification branch uses Focal Loss with a class balance factor set to 0.25 and a focusing parameter γ set to 2; the regression branch uses DIoU loss combined with a CIoU improvement term.

[0064] Based on the general feature processing of preprocessed images containing multi-scale information using neural implicit fields, 6D pose estimation results are obtained. This step first performs neural implicit field modeling, including a geometric field and an appearance field. The geometric field takes feature vectors and 3D spatial point coordinates as input, and outputs signed distance field (SDF) values ​​through a multilayer perceptron (MLP) network to describe the geometry of the target surface. The MLP network contains four hidden layers, each with 256 neurons, using the ReLU activation function. The appearance field combines 3D point normal information and predicts the RGB color of the corresponding points through another MLP network to extract illumination-invariant features. The appearance field MLP network contains three hidden layers, each with 128 neurons, using the ReLU activation function. Then, pose hypothesis generation and optimization are performed. First, based on the prior knowledge of the detection box center and target size, an initial pose set including random rotations and translations is generated, with the initial pose count set to 64. A virtual view is synthesized through neural implicit field rendering, and feature matching is calculated with the input image. Low-confidence hypotheses are filtered out, and the top 16 best hypotheses are retained. Finally, the gradient descent algorithm was used to iteratively adjust the pose parameters with the SDF value and image reprojection error as the optimization objectives. The number of iterations was set to 50, the initial learning rate was 0.01, and the cosine annealing strategy was used for adjustment.

[0065] This step incorporates the FoundationPose network structure. In the data generation section, based on the detection results, a language model (such as GPT-3) is used to generate a description of the target's appearance. This description is then input into a diffusion model (such as StableDiffusion) to enhance the appearance of the original model, generating models with diverse appearances. From the perspective of the physics engine and path tracing, the detected target location information is used to precisely locate and crop the rendering area, rendering only the region of interest containing the target. Based on the target's geometric features and material properties, the simulation parameters in the physics engine are optimized to make the rendered image more realistic. In the model-free setting of Neural Object Modeling, the target bounding box information provided by the detection module is used to preprocess the reference image, enlarging and enhancing the target region in the reference image. The resulting image features are then input into the geometry and appearance networks. The structures of the geometry and appearance networks are fine-tuned, and a RALA module is introduced to enhance the network's ability to extract target geometric and appearance features. When training the geometry network and appearance network, a multi-task learning approach is adopted, incorporating auxiliary information from the object detection task as additional supervision signals into the network training process, with the weight coefficient λ set to 0.1.

[0066] In the pose hypothesis generation stage, the global pose sampling method based on FoundationPose is optimized by combining detected target semantic information. For symmetrical targets, repeated sampling in symmetrical directions is reduced; for targets with specific functional structures, sampling density in functionally relevant directions is increased. In the encoder part, the RALA module is used to process the input features and enhance feature representation. When updating rotation and translation through the Transformer encoder, target scale information output by the detection module is introduced to adjust the pose update parameters. During the update process, a hierarchical contrastive network method is used to compare and filter pose hypotheses at different levels, prioritizing the retention of hypotheses that are closer to the actual target pose. In the Pose Ranking Encoder, the RALA module is introduced to enhance the extraction and comparison capabilities of features from different pose hypotheses. When calculating the pose ranking, the ranking of pose hypotheses is weighted and adjusted by combining target confidence information provided by the detection module. When processing multiple pose hypotheses, a multi-hypothesis fusion strategy is adopted to fuse the features of different pose hypotheses, and then feature interaction and filtering are performed through a self-attention mechanism.

[0067] By coordinating 6D pose estimation results through a multi-task loss function, the parameters of the shared feature extraction module for both object detection and pose estimation are optimized synchronously through backpropagation. This step designs a comprehensive loss function comprising two parts: object detection loss and pose estimation loss. The object detection loss includes classification loss (Focal Loss) and bounding box regression loss (CIoU Loss), while the pose estimation loss includes rotation loss (L2 loss based on quaternion representation) and translation loss (L1 loss). The weights for classification loss, bounding box regression loss, rotation loss, and translation loss are all set to 1.0. During training, the parameters of both the shared feature extraction module and the 6D pose estimation module are updated simultaneously, allowing the two tasks to mutually promote each other and improve overall performance. The Adam optimizer is used for parameter updates, with an initial learning rate of 0.001, a weight decay coefficient of 0.0001, a batch size of 16, and 100 training epochs.

[0068] Example 4 like Figure 2 As shown, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the pose estimation method for a live-line working robot in embodiment three.

[0069] The electronic device is equipped with a high-performance CPU or GPU with a clock speed of at least 2.5GHz, at least 8 cores, and supports CUDA acceleration. Memory includes RAM and storage space, with at least 16GB of RAM and at least 256GB of storage space. The device also features a dedicated AI acceleration chip to accelerate the inference process of deep learning models. It connects to the RGB-D camera of the power distribution network live-line working robot via a high-speed interface to receive image data in real time. The device runs on Linux or Windows operating systems and supports deep learning frameworks such as PyTorch or TensorFlow.

[0070] The computer program of the electronic device implements all the functions of the pose estimation method for the live-line working robot in the power distribution network described in Example 3. The program is modularly designed, including an image preprocessing module, a feature extraction module, a target detection module, a pose estimation module, and a joint optimization module. Data exchange between modules is achieved through standardized interfaces, ensuring the system's scalability and maintainability. The program supports multi-threaded parallel processing, fully utilizing the computing power of multi-core processors to improve the system's real-time performance. The program also implements a data caching mechanism to cache frequently used features and model parameters, reducing redundant calculations and improving processing efficiency.

[0071] Example 5 A computer-readable storage medium storing a computer program that, when executed by a processor, implements the pose estimation method for a live-line working robot in a power distribution network as described in Embodiment 3.

[0072] The computer-readable storage medium can be a solid-state drive, a hard disk drive, an optical disc, a USB flash drive, or other non-volatile storage devices. The computer program stored in the storage medium contains the complete implementation code of the pose estimation method for the live-line working robot in the power distribution network described in Embodiment 3. The program adopts a modular design with clear interfaces between functional modules, facilitating maintenance and upgrades. The program also includes necessary configuration files and pre-trained model files to ensure rapid system deployment and use.

[0073] The computer program is optimized for different hardware platforms, supporting both CPU and GPU operating modes. In GPU mode, the program fully utilizes the CUDA acceleration library to improve the inference speed of deep learning models. The program also implements an automatic adaptation mechanism that automatically adjusts parameters based on the hardware configuration of the operating environment, ensuring optimal performance on devices with varying configurations.

[0074] Example 6 A computer program product, the computer program product including computer instructions, the computer instructions instructing a computer to execute the pose estimation method for a live-line working robot in a power distribution network as described in Embodiment 3.

[0075] The computer program product is provided as a software package, containing the complete implementation code and necessary dependency libraries for the pose estimation method of the live-line working robot in the power distribution network described in Embodiment 3. The software package provides a graphical user interface for user operation and monitoring of the system's operating status. The user interface consists of three parts: an image display area, a parameter configuration area, and a result display area. The image display area displays the original RGB-D image, the preprocessed image, and visualizations of the detection and pose estimation results in real time. The parameter configuration area allows users to adjust various system parameters, such as preprocessing parameters, detection thresholds, and pose estimation accuracy requirements. The result display area displays the detected target category, location, and 6D pose information, as well as the system's operating status and performance indicators.

[0076] The software package also provides an API interface, allowing other systems to call the system's functions through standardized interfaces. The API interface supports both RESTful style and WebSocket calling methods, facilitating integration with different types of systems. The package also includes detailed documentation and sample code to help developers quickly understand and use the system's functions.

[0077] It should be noted that Embodiment 1, Embodiment 2, Embodiment 3, Embodiment 4, Embodiment 5, and Embodiment 6 are all types of pose estimation systems for live-line working robots in power distribution networks.

[0078] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0079] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0080] This application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, readable storage media, optical storage, etc.) containing computer-usable program code.

[0081] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0082] Obviously, the described embodiments are only some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort should fall within the scope of protection of this application.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit them. Although this application has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of this application. Any modifications or equivalent substitutions that do not depart from the spirit and scope of this application should be covered within the protection scope of the claims of this application.

Claims

1. A pose estimation system for a live-line working robot in a power distribution network, characterized in that, include: The input preprocessing module is used to perform noise reduction and enhancement processing on the RGB-D images acquired by the live-line working robot of the power distribution network to obtain preprocessed images; A shared feature extraction module is used to extract general features containing multi-scale information in preprocessed images based on a lightweight convolutional architecture. The 6D pose estimation module is used to process general features containing multi-scale information in the preprocessed image based on the neural implicit field to obtain the 6D pose estimation result.

2. The pose estimation system for a live-line working robot in a power distribution network according to claim 1, characterized in that, The input preprocessing module includes: The image denoising unit is used to suppress salt-and-pepper noise in power distribution scenarios by using adaptive median filtering for RGB-D images, and to eliminate the influence of uneven illumination by combining homomorphic filtering to obtain the denoised image; The depth map optimization unit is used to smooth the depth values ​​of the denoised image through bilateral filtering, fill in invalid pixels, and obtain a preprocessed image.

3. The pose estimation system for a live-line working robot in a power distribution network according to claim 1, characterized in that, The shared feature extraction module is specifically used for: Multi-stage feature extraction: shallow features extract basic information; mid-level features introduce the bottleneck structure Bottleneck and combine it with dilated convolution to capture medium-sized targets; Deep features: Employs a cross-stage local connectivity structure to output high-level features; Feature enhancement: Embedding RALA modules in the backbone network.

4. The pose estimation system for a live-line working robot in a power distribution network according to claim 1, characterized in that, The shared feature extraction module uses YOLOv12 as the core architecture for object detection. In YOLOv12, RALA is embedded between the convolutional layers in the Backbone part, residual connections are introduced in the downsampling module C3k2, RALA is used in the Neck part, and a decoupled design is adopted in the Head part, along with a specific loss function. Specifically, in the Backbone part, the input preprocessed image first goes through multiple convolutional layers for preliminary feature extraction; in the original YOLOv12 Backbone, a rank-enhanced linear attention module is embedded after some convolutional layers. The Neck section fuses feature maps of different scales. Specifically, it uses the RALA attention mechanism in the A2C2f module. For the feature maps in the A2C2f module, RALA is applied to each block region according to a block-based strategy, with a computational complexity of O(n log n). , where N is the total number of pixels in the feature map; In terms of upsampling and stitching, the Upsample module is used to upsample low-resolution feature maps to high-resolution ones. It adopts an adaptive upsampling method, which dynamically adjusts the upsampling weights according to the content of the feature map and the distribution of the target. The head section of the detection head adopts a decoupled design, divided into classification branch and regression branch; The classification branch uses FocalLoss, calculated as follows: ,in As a class balance factor, To focus on the parameters, the regression branch uses DIoU loss combined with CIoU improvement terms. ,in .

5. The pose estimation system for a live-line working robot in a power distribution network according to claim 4, characterized in that, The rank-enhanced linear attention module is used for: The rank enhancement mechanism in the KV buffer fuses detection semantics and spatial features to generate context-aware global queries. : Where c is the detection category label, which is mapped to a 64-dimensional vector through a learnable embedding layer; then the weight modulation is calculated. , ,in It is a nonlinear kernel function; in the rank recovery of the output features, a channel interaction matrix is ​​introduced. Features of the original token Perform linear transformation , ; The downsampling module C3k2 is used to perform downsampling operations. In the C3k2 module, the input features are divided into two paths. One path goes through a series of convolution operations, and the other path is directly connected to the output of the module. Finally, the results of the two paths are added together.

6. The pose estimation system for a live-line working robot in a power distribution network according to claim 4, characterized in that, The target detection branch includes: The multi-scale feature fusion unit is used to upsample and concatenate the features output from different levels of the backbone network through the neck network to generate a fusion feature map containing rich details and semantics. An anchor box mechanism is applied to the feature map to predict the class probability of the target and the coordinates of the 2D detection box. The loss function unit uses CIoU loss to calculate the regression error of the detection box.

7. The pose estimation system for a live-line working robot in a power distribution network according to claim 1, characterized in that, The 6D pose estimation module is used for: 1) Neural implicit field modeling, including: geometric field, which takes feature vectors and 3D spatial point coordinates as input and outputs signed distance field values ​​to describe the geometry of the target surface; appearance field, which combines 3D point normal information to predict the RGB color of the corresponding point and extract illumination invariant features. 2) Attitude hypothesis generation and optimization, including: Initial assumptions are made based on prior knowledge of the detection box center and target size, generating an initial pose set that includes random rotations and translations; Hypothesis filtering: A virtual view is synthesized through neural implicit field rendering, and its feature matching degree is calculated with the input image to filter out low-confidence hypotheses; Iterative optimization is performed using the gradient descent algorithm, with the SDF value and image reprojection error as the optimization targets, to iteratively adjust the pose parameters.

8. The pose estimation system for a live-line working robot in a power distribution network according to claim 1, characterized in that, The 6D pose estimation module incorporates the FoundationPose network structure. In the data generation part, synthetic data is generated by combining detection results with language models and diffusion models, and the rendering area and physical simulation parameters are optimized. In the neural object modeling part, reference images are preprocessed with detection bounding boxes, the network structure is fine-tuned and RALA is introduced, and multi-task training is performed using detection auxiliary information. In the pose hypothesis generation part, the sampling strategy is optimized based on the detection semantic information, the feature expression is enhanced by RALA, and the pose update is improved by combining the detection scale information; In the pose selection part, RALA is introduced into the encoder, and the pose ranking is adjusted in combination with the detection confidence to perform multiple hypothesis fusion and screening.

9. The pose estimation system for a live-line working robot in a power distribution network according to claim 1, characterized in that, The 6D pose estimation module specifically includes: In FoundationPose, based on the target category information output by the detection module, an LLM is used to generate a description of the target's appearance. This description is then input into the diffusion model to enhance the appearance of the original model, generating models with diverse appearances. From the perspective of the physics engine and path tracing, combined with the detected target location information, the rendering area is precisely located and cropped. Only the region of interest containing the target is rendered. Based on the target's geometric features and material properties, the simulation parameters in the physics engine are optimized to make the rendered image more realistic. Let the target's geometric feature parameter be G, the material property parameter be M, and the rendering area parameter be R. The physics engine parameters are adjusted using the function f(G,M,R). In the model-free setting of Neural Object Modeling, based on the FoundationPose architecture, a neural RGBD object field is constructed using a small number of reference images. The reference images are preprocessed using the target bounding box information provided by the detection module, and the target regions in the reference images are enlarged and enhanced. The resulting image features are then input into the geometry network and the appearance network. The structure of the geometry network and the appearance network is fine-tuned, and a RALA module is introduced into the network. In a geometric network, let the input features be... The output after RALA processing Through a series of convolution operations and RALA operation The calculation process is as follows: This enhances the network's ability to extract the geometric features of the target. In the appearance network, let the input features be... The output is ,go through and Operation, i.e. This improves the ability to extract target appearance features, thereby enhancing the modeling accuracy of the neural implicit field. When training the geometry and appearance networks, a multi-task learning approach is adopted, incorporating auxiliary information from the target detection task as additional supervision signals into the network training process. Let the auxiliary information for the detection task be... The geometric network loss function is The appearance network loss function is The total loss function is ,in These are the weighting coefficients; In the pose hypothesis generation stage, the global pose sampling method based on FoundationPose is optimized by combining the detected target semantic information to generate pose hypotheses. For symmetrical targets, reduce repeated sampling in the symmetrical direction; for targets with specific functional structures, increase the sampling density in their functionally relevant directions; in the encoder section, the RALA module is used to process the input features; let the input features be... The output after RALA processing is ,pass Enhanced features: When updating rotation and translation via the Transformer encoder, the target scale information S output by the detection module is introduced to adjust the pose update parameters; let the rotation parameter be r and the translation parameter be t, then r and t are adjusted using the function g(S). , During the update process, a hierarchical contrastive network method is used to compare and filter pose assumptions at different levels, prioritizing the retention of assumptions that are closer to the actual pose of the target. In the Pose Ranking Encoder, the structure of FoundationPose is borrowed, and a RALA module is introduced to enhance the extraction and comparison capabilities of features from different pose assumptions. RALA is applied in the convolutional residual blocks and self-attention layers of the encoder. Let the input pose assumption features be... The output after RALA processing is ,pass This allows the model to better capture the differences between different pose assumptions; when calculating the pose ranking, the target confidence information C provided by the detection module is used to weight and adjust the ranking of the pose assumptions; let the initial ranking score of the pose assumption be... The weighted score is Adjusted by the function h(C), i.e. ; When processing multiple pose hypotheses, a multi-hypothesis fusion strategy is adopted to fuse the features of different pose hypotheses, and then feature interaction and filtering are performed through a self-attention mechanism. A constraint based on the target detection results is also introduced for pose hypotheses that are obviously inconsistent with the detected target position and appearance.

10. The pose estimation system for a live-line working robot in a power distribution network according to claim 1, characterized in that, Also includes: The joint optimization module is used to coordinate the 6D pose estimation results through a multi-task loss function, detect the shared feature extraction module parameters with the pose estimation, and synchronously optimize the 6D pose estimation module through backpropagation.

11. A pose estimation method for a live-line working robot in a power distribution network, based on the pose estimation system for a live-line working robot in a power distribution network as described in any one of claims 1 to 10, characterized in that, include: The RGB-D images acquired by the live-line working robot in the power distribution network are denoised and enhanced to obtain pre-processed images; Based on a lightweight convolutional architecture, general features containing multi-scale information in preprocessed images are extracted. Based on the general feature processing of preprocessed images containing multi-scale information using neural implicit fields, 6D pose estimation results are obtained.

12. The pose estimation method for a live-line working robot in a power distribution network according to claim 11, characterized in that, Also includes: By coordinating the 6D pose estimation results through a multi-task loss function, the detection and pose estimation share the same feature extraction module parameters, and the 6D pose estimation results are optimized synchronously through backpropagation.

13. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the pose estimation method for a live-line working robot of claim 11 or 12.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the pose estimation method for a live-line working robot in a power distribution network as described in claim 11 or 12.

15. A computer program product, the computer program product comprising computer instructions, characterized in that, The computer instructions instruct the computer to execute the pose estimation method for the live-line working robot of the power distribution network as described in claim 11 or 12.

Images