Live-line robot work target positioning method and system

Abstract

The embodiment of the application provides a working target positioning method and system of a live working robot, and belongs to the technical field of distribution network live working. The method comprises the following steps: acquiring image information containing a working target, and pre-processing the image information to obtain a pre-processed image; taking the pre-processed image as an input image of a pre-optimized VGG-UNet model, performing working target identification, and fitting a center line of the identified working target; obtaining a disparity map based on a semi-global stereo matching method and the fitted center line; and solving and obtaining the depth and position of the working target based on the disparity map. The scheme has high accuracy in working target identification and strong anti-interference capability, and provides an identification basis for the precise working of the working robot.

Description

Technical Field

[0001] This invention relates to the field of live-line working technology in power distribution networks, and more specifically to a target positioning method and a target positioning system for a live-line working robot. Background Technology

[0002] Live-line work is a crucial means to improve power supply reliability and enhance power service levels. However, the compact structure and small safety distances of distribution network equipment, coupled with the high-altitude and high-voltage working environment, pose a significant threat to the personal safety of construction workers. Furthermore, live-line work in distribution networks is extremely demanding, and the quality of work varies greatly from person to person. Robots, on the other hand, can remotely perform tasks in place of operators. Leveraging this characteristic, robots can be used in power operation scenarios to replace manual labor in live-line maintenance. Currently, the mainstream methods for robot target recognition include LiDAR, structured light cameras, and binocular vision. LiDAR is expensive and generates sparse point clouds, making it difficult to generate point clouds for thin, weakly textured conductors. Structured light cameras are easily affected by sunlight, resulting in poor outdoor performance. Binocular vision, with its rich semantic information and high flexibility, performs well in distribution network scenarios and has become the primary means for distribution network robots to identify and locate thin, weakly textured targets such as conductors. However, due to severe sunlight interference during distribution network operations, image matching for thin, weakly textured targets such as conductors is difficult, and the point cloud quality generated by binocular vision is poor, making conductor identification and localization challenging. Therefore, target recognition under strong outdoor light interference is one of the biggest challenges for distribution network robots.

[0003] In the process of target recognition based on binocular cameras, the key lies in stereo matching. This involves using a binocular camera to simulate human vision observing a scene from two different angles. By matching corresponding pixels in the two images, calculating the disparity, and then using triangulation to calculate the depth information of the corresponding pixels, the three-dimensional information of the scene is obtained. Existing stereo matching algorithms are broadly divided into two categories: intensity-based methods, which are highly sensitive to noise caused by changes in lighting or camera lens shake; and feature-based methods, which rely on finding sparse features such as Harris corners, SURF features, and edge features. However, due to the instability of features, feature-based methods have certain limitations. To address the technical problem of difficulty in stereo matching under outdoor sunlight interference in existing binocular vision systems for live-line working robots, leading to difficulties in target recognition and low positioning accuracy, a new target localization method for live-line working robots is needed. Summary of the Invention

[0004] The purpose of this invention is to provide a target localization method and system for live-line working robots, so as to at least solve the technical problems of difficulty in stereo matching of binocular vision of existing power distribution network live-line working robots under outdoor sunlight interference, resulting in difficulty in target identification and low positioning accuracy.

[0005] To achieve the above objectives, the first aspect of the present invention provides a method for locating a work target for a live-line working robot. The method includes: acquiring image information containing the work target and preprocessing the image information to obtain a preprocessed image; using the preprocessed image as input to a pre-optimized VGG-UNet model to identify the work target and fitting the centerline of the identified work target; obtaining a disparity map based on a semi-global stereo matching method and the fitted centerline; and solving for the depth and position of the work target based on the disparity map.

[0006] Optionally, the preprocessing of the image information includes: performing image correction processing and image enhancement processing.

[0007] Optionally, image correction includes: obtaining camera calibration parameters and distortion parameters; wherein the camera is a stereo camera; the camera calibration parameters include preset internal parameters and preset external parameters of the camera; calculating an image projection matrix based on the camera calibration parameters and distortion parameters; correcting non-coplanar line aligned stereo images to coplanar line aligned images, and performing a one-dimensional search on the epipolar line to match corresponding points in the two images, so that the row pixels of the stereo images are aligned one by one.

[0008] Optionally, image enhancement processing includes: extracting an image enhancement curve from the image after image correction, the curve being represented as:

[0009] LE(I(x,y);α)=I(x,y)+αI(x,y)(1-I(x,y))

[0010] Where LE represents the pixel enhancement result; α represents the preset image enhancement coefficient; I(x,y) represents the light intensity value at pixel coordinates (x,y); the light intensity value at each pixel coordinate point is mapped multiple times through the image enhancement curve to obtain the light intensity value after each mapping; wherein, the conversion formula for the light intensity value after each mapping is:

[0011] LE n (I(x,y);α n ) = I n-1 (x,y)+A n I n-1 (x,y)(1-I n-1 (x,y))

[0012] Among them, LE nThis represents the pixel enhancement result of the nth mapping iteration; α n I represents the image enhancement coefficient at the nth mapping; n-1 (x,y) represents the light intensity value after the (n-1)th mapping; A n A pre-introduced image enhancement parameter matrix of the same size as the input image is used; the optimal image enhancement parameter matrix is ​​found based on the Zero-DCE network, and the enhanced light intensity value is obtained based on the optimal image enhancement parameter matrix; the corresponding enhanced image is obtained based on the enhanced light intensity value.

[0013] Optionally, the method further includes: constructing a pre-optimized VGG-UNet model, including: constructing a VGG16 network and a UNet network respectively, and combining the two models to obtain a VGG-UNet network; wherein, combining the two models includes: removing the fully connected layers in the VGG16 network and replacing them with the enhanced feature extraction network of the UNet network; collecting historical image information containing the task target, and training a VGG-UNet model for image segmentation of the task target in the VGG-UNet network based on the historical image information.

[0014] Optionally, the historical image information containing the task target includes: historical image information of the task target under different sunlight angles, different weather conditions, different shooting angles, and different shooting postures; the step of training a VGG-UNet model for image segmentation of the task target in the VGG-UNet network based on the historical image information includes: performing image enhancement operations on the historical image information, then annotating the image after image enhancement to obtain an annotated dataset; and importing the annotated dataset into the VGG-UNet network for training to obtain the VGG-UNet model.

[0015] Optionally, the step of using the preprocessed image as the input image of the pre-optimized VGG-UNet model for target recognition and fitting the centerline of the recognized target includes: importing the preprocessed image into the VGG-UNet model for target segmentation and highlighting the segmented target regions; and fitting the centerline of the highlighted regions using the image center moments to obtain the centerline of the target in the image.

[0016] Optionally, the step of obtaining the depth and position of the target based on the disparity map includes: obtaining the center distance, focal length, and disparity of corresponding left and right points of the two cameras; calculating the depth value of each pixel based on the center distance, focal length, and disparity of corresponding left and right points of the two cameras; calculating the average depth value of each pixel of the target and using the average value as the distance from the target to the camera; and determining the position of the target based on the distance.

[0017] A second aspect of the present invention provides a target localization system for a live-line working robot. The system includes: a data acquisition unit, configured to acquire image information containing the target and preprocess the image information to obtain a preprocessed image; a processing unit, configured to: use the preprocessed image as input to a pre-optimized VGG-UNet model, perform target recognition, and fit the centerline of the recognized target; obtain a disparity map based on a semi-global stereo matching method and the fitted centerline; and a localization unit, configured to solve for the depth and position of the target based on the disparity map.

[0018] On the other hand, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described method for locating the target of a live-lined robot.

[0019] Through the above technical solution, the present invention first performs correction and enhancement processing on the acquired target image information, resulting in higher accuracy for subsequent feature extraction based on the processed image. Then, centerline fitting is performed on the identified target image, and a disparity map is obtained based on the semi-global stereo matching method and the fitted centerline. Finally, the specific depth and position of the target are obtained based on this disparity map. The present invention provides high accuracy and strong anti-interference capability for target recognition, providing a foundation for precise operation by robots.

[0020] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0022] Figure 1 This is a flowchart of the steps of a live-line working robot target localization method according to one embodiment of the present invention;

[0023] Figure 2 This is a system structure diagram of a target positioning system for a live-line working robot provided in one embodiment of the present invention. Detailed Implementation

[0024] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0025] Live-line work is a crucial means to improve power supply reliability and enhance power service levels. However, the compact structure and small safety distances of distribution network equipment, coupled with the high-altitude and high-voltage working environment, pose a significant threat to the personal safety of construction workers. Furthermore, live-line work in distribution networks is extremely demanding, and the quality of work varies greatly from person to person. Robots, on the other hand, can remotely perform tasks in place of operators. Leveraging this characteristic, robots can be used in power operation scenarios to replace manual labor in live-line maintenance. Currently, the mainstream methods for robot target recognition include LiDAR, structured light cameras, and binocular vision. LiDAR is expensive and generates sparse point clouds, making it difficult to generate point clouds for thin, weakly textured conductors. Structured light cameras are easily affected by sunlight, resulting in poor outdoor performance. Binocular vision, with its rich semantic information and high flexibility, performs well in distribution network scenarios and has become the primary means for distribution network robots to identify and locate thin, weakly textured targets such as conductors. However, due to severe sunlight interference during distribution network operations, image matching for thin, weakly textured targets such as conductors is difficult, and the point cloud quality generated by binocular vision is poor, making conductor identification and localization challenging. Therefore, target recognition under strong outdoor light interference is one of the biggest challenges for distribution network robots.

[0026] In target recognition based on binocular cameras, the key lies in stereo matching. This involves using a binocular camera to simulate human vision observing a scene from two different angles. By matching corresponding pixels in the two images, calculating the disparity, and then using triangulation to calculate the depth information of the corresponding pixels, the three-dimensional information of the scene is obtained. Existing stereo matching algorithms can be broadly divided into two categories: intensity-based methods, which are highly sensitive to noise caused by changes in lighting or camera lens shake; and feature-based methods, which rely on finding sparse features such as Harris corners, SURF features, and edge features. However, due to the instability of features, feature-based methods have certain limitations.

[0027] This invention addresses the technical problem of existing live-line working robots in power distribution networks facing difficulties in stereo matching under outdoor sunlight interference, resulting in challenges in target identification and low positioning accuracy. It proposes a target localization method for live-line working robots to reduce the error rate in identifying and locating targets under strong outdoor light interference.

[0028] Figure 1 This is a flowchart of a method for locating a work target for a live-line working robot according to one embodiment of the present invention. Figure 1 As shown, this invention provides a method for locating a work target for a live-line working robot, the method comprising:

[0029] Step S10: Obtain image information containing the task target, and preprocess the image information to obtain a preprocessed image.

[0030] Specifically, the present invention uses images acquired by a binocular camera for target localization through corresponding image processing. To achieve target localization within binocular images, both the accuracy of target recognition and subsequent localization must be guaranteed. Because images directly acquired by the camera are limited by the acquisition environment and camera limitations, they contain numerous noise and interference components. These interference components need to be removed, and the images enhanced to ensure the accuracy of subsequent target recognition.

[0031] Furthermore, after the binocular camera completes image acquisition, the images need to be preprocessed. This preprocessing includes image correction and image enhancement. Image correction includes obtaining camera calibration parameters and distortion parameters; wherein the camera is a binocular camera; the camera calibration parameters include preset internal and external parameters; calculating the image projection matrix based on the camera calibration parameters and distortion parameters; correcting non-coplanar line-aligned binocular images to coplanar line alignment, and performing a one-dimensional search on the epipolar lines to match corresponding points in the two images, thus aligning the row pixels of the binocular images one-to-one.

[0032] In one possible implementation, the Zhang Zhengyou calibration method is used to calibrate the stereo camera. To reduce experimental errors, a 12*9 grid series calibration plate printed on a GP340 float glass substrate is used, with a grid side length of 25mm and a side length accuracy of 0.01mm. To solve for the camera's intrinsic and extrinsic parameters, it is only necessary to photograph the same checkerboard calibration plate from different angles using the stereo camera. The camera's intrinsic and extrinsic parameters can be directly obtained by the correspondence between the corner points on the calibration plate and the corner points in the images obtained by the camera. This can be represented as:

[0033]

[0034] Where A is the camera's intrinsic parameter matrix; R is the rotation matrix of the extrinsic parameters; and T is the translation vector. f is the focal length; dx and dy are the width and height of the pixel, respectively; γ represents the deviation of the pixel in the x and y directions; (μ0, ν0) is the reference point.

[0035] Furthermore, when solving for the camera distortion parameters, multiple chessboard images are acquired, and the corresponding intrinsic and extrinsic parameters are calculated for each. Assuming the initial camera distortion is zero, an optimization function is given, and the camera distortion parameters corresponding to the minimum output cost of the optimization function are obtained. The image projection matrix is ​​calculated using the camera intrinsic, extrinsic, and distortion parameters obtained from camera calibration. The binocular images are projected onto the focal plane located at a distance f from the optical center of the binocular camera along the optical axis. Corresponding points in the two images are matched to complete image correction.

[0036] Furthermore, image enhancement processing includes: extracting an image enhancement curve from the image after image correction, which is represented as:

[0037] LE(I(x,y);α)=I(x,y)+αI(x,y)(1-I(x,y))

[0038] Where LE represents the pixel enhancement result; α represents the preset image enhancement coefficient; I(x,y) represents the light intensity value at pixel coordinates (x,y); the processed light intensity value is then mapped again through the enhancement curve, thereby enhancing the nonlinearity of the curve and enabling it to better cope with various light intensities and extreme lighting conditions. At this point, the light intensity value is transformed into:

[0039] LE n (I(x,y);α n ) = I n-1 (x,y)+A n I n-1 (x,y)(1-I n-1 (x,y))

[0040] Among them, LE n This represents the pixel enhancement result of the nth mapping iteration; α n I represents the image enhancement coefficient at the nth mapping; n-1 (x,y) represents the light intensity value after the (n-1)th mapping.

[0041] Furthermore, an enhancement parameter matrix A is introduced, and the light intensity value at each pixel coordinate point is processed by the aforementioned graph.

[0042] n

[0043] The enhancement curve is mapped multiple times to obtain the light intensity value after each mapping round; the conversion formula for the light intensity value after each mapping round is as follows:

[0044] LE n (I(x,y);α n ) = I n-1 (x,y)+A n I n-1 (x,y)(1-I n-1 (x,y))

[0045] Among them, LE n This represents the pixel enhancement result of the nth mapping iteration; α n I represents the image enhancement coefficient at the nth mapping; n-1 (x,y) represents the light intensity value after the (n-1)th mapping; A nA pre-introduced image enhancement parameter matrix of the same size as the input image is used; the optimal image enhancement parameter matrix is ​​found based on the Zero-DCE network, and the enhanced light intensity value is obtained based on the optimal image enhancement parameter matrix; the corresponding enhanced image is obtained based on the enhanced light intensity value.

[0046] In one possible implementation, the ZED-2i binocular camera is used as the visual perception camera of the power distribution network live-line working robot. The ZED-2i is used to acquire RGB images, and the RGB images are processed by the Zero-DCE (Zero Low Light Enhancement) algorithm without reference data to obtain the enhanced image, reducing the interference caused by direct sunlight or backlight to the image.

[0047] Step S20: Use the preprocessed image as the input image of the pre-optimized VGG-UNet model to perform target recognition and fit the centerline of the recognized target.

[0048] Specifically, the present invention utilizes an optimized VGG-UNet model for target identification and labeling. Therefore, before target identification, the method further includes constructing a pre-optimized VGG-UNet model, comprising: constructing a VGG16 network and a UNet network respectively, and combining the two models to obtain the VGG-UNet network; wherein, combining the two models includes: removing fully connected layers from the VGG16 network and replacing them with enhanced feature extraction networks from the UNet network; collecting historical image information containing the target, and training a VGG-UNet model for image segmentation of the target based on the historical image information in the VGG-UNet network.

[0049] In this embodiment of the invention, VGG16 includes 13 convolutional layers, 3 fully connected layers, and 5 pooling layers. Due to its large parameter list, it has a high fitting capability and is often used for image object classification and localization. However, precisely because of its large parameter list, its training time is too long, parameter tuning is difficult, and it requires a large storage capacity. This invention aims to retain the strong fitting capability of the VGG16 network while avoiding its low training efficiency, and therefore incorporates a UNet network.

[0050] Furthermore, semantic segmentation is an important branch of image processing and machine vision. Unlike classification tasks, semantic segmentation requires determining the category of each pixel in an image for accurate segmentation. Semantic segmentation currently has wide applications in autonomous driving, automatic image matting, and medical imaging. UNet is arguably the most commonly used and simplest segmentation model; it is simple, efficient, easy to understand, easy to build, and can be trained on small datasets. This invention discards the three resource-intensive fully connected layers of VGG16 and replaces them with an enhanced feature extraction network of UNet, retaining the first four downsampling layers of the VGG16 network, making it deeper than the traditional UNet network. Using UNet skip connections preserves low-level features, avoiding the potential loss of low-level features during downsampling, and connects them to high-level features after upsampling.

[0051] Furthermore, after completing the VGG-UNet model construction, image recognition can be performed based on this model. The step of using the preprocessed image as input to the pre-optimized VGG-UNet model for target recognition and fitting the centerline of the recognized target includes: importing the preprocessed image into the VGG-UNet model, segmenting the target, and highlighting the segmented target regions; and fitting the centerline of the highlighted regions using the image's central moments to obtain the centerline of the target in the image.

[0052] Preferably, the historical image information containing the task target includes: historical image information of the task target under different sunlight angles, different weather conditions, different shooting angles, and different shooting postures; the step of training a VGG-UNet model for image segmentation of the task target in the VGG-UNet network based on the historical image information includes: performing image enhancement operations on the historical image information, then annotating the image after image enhancement to obtain an annotated dataset; and importing the annotated dataset into the VGG-UNet network for training to obtain the VGG-UNet model.

[0053] In another possible implementation, when historical images of the target task are lacking, a target task can be selected, experimental shooting can be conducted, different shooting conditions can be created, images can be acquired based on specific conditions, and then this acquired image information can be used as historical image information. Taking the target task as a guideline as an example, the specific steps for model construction and subsequent image segmentation can be represented as follows:

[0054] 1) Collect target object dataset. Depending on the type of target object to be captured, collect images of the target object from different sunlight angles, different weather conditions, different shooting angles, and different poses;

[0055] 2) Data augmentation, which expands the dataset through random cropping, brightness transformation, adding noise, flipping, rotating, etc.

[0056] 3) Label the dataset: Use image annotation tools to label the collected images;

[0057] 4) VGG-UNet model optimization;

[0058] 5) Model transfer learning training: Using the transfer learning method, the self-made dataset is loaded into the optimized network model, and the model pre-trained using the VOC dataset is loaded at the same time to improve model convergence. Through parameter optimization, the model is iteratively trained to generate object detection and segmentation models.

[0059] 6) For target detection and segmentation, the image enhanced in step S10 is input into the VGG-UNet model. The target is segmented by identifying the target object category and target location in the live-line operation of the distribution network through pattern recognition, and the target area is highlighted in blue.

[0060] 7) Determine the centerline of the target, and perform centerline fitting on the segmented regions using the image center moments to calculate the centerline of the target conductor in the binocular image.

[0061] Step S30: Obtain the disparity map based on the semi-global stereo matching method and the fitted centerline.

[0062] Specifically, the semi-global stereo matching method is a matching algorithm that lies between local and global matching. It effectively combines the advantages and disadvantages of both methods, achieving a good balance between accuracy and efficiency. It is now widely used in many commercial software applications, particularly in aerial remote sensing imagery. Its core components include the cost calculation of hierarchical mutual information, multi-directional one-dimensional approximation of two-dimensional cost aggregation, and other refined post-processing parts, such as left-right consistency processing and sub-pixel refinement. Because of its left-right consistency processing capability, it is often used to determine the disparity map of binocular images. Based on this characteristic, the present invention applies it to a specific solution.

[0063] Step S40: Obtain the depth and position of the target based on the disparity map.

[0064] Specifically, the step of obtaining the depth and position of the target based on the disparity map includes: obtaining the center distance, focal length, and disparity of corresponding left and right points of the two cameras; calculating the depth value of each pixel based on the center distance, focal length, and disparity of corresponding left and right points of the two cameras; calculating the average depth value of each pixel of the target and using the average value as the distance from the target to the camera; and determining the position of the target based on the distance.

[0065] In one possible implementation, based on the ranging principle of a stereo camera, the depth Z depends only on three parameters: the distance B between the centers of the two cameras, the focal length f of the cameras, and the disparity d between corresponding points on the left and right. B and f are constants for the selected stereo camera; therefore, after obtaining the disparity map, the value of d for each point is known, and thus the depth Z value for each point can be derived. Therefore, from the disparity map in step S30, the depth Z of each pixel in the ROI in the image can be calculated. The average depth Z of the pixels corresponding to the target guideline is calculated, and this average value is used as the distance from the target object to the camera. The guideline is fitted with a straight line using the Random Sampling Consistent Arithmetic (RANSAC) method to obtain the straight line equation of the guideline, thus completing the localization of the target guideline.

[0066] In this embodiment of the invention, the image enhancement part uses Zero-DCE for low-light enhancement, improving images under extreme lighting conditions such as direct sunlight and backlighting. The enhanced image also significantly improves the performance of conductor segmentation, enhancing the adaptability of the live-line working robot to the working environment. In image segmentation, a VGG-UNet-based target recognition and segmentation method is proposed to segment the target region (ROI). Then, the image center moment is used to fit the midline of the segmented region. By replacing direct matching of the conductor image with matching the conductor centerline, the robustness and stability of the algorithm are greatly enhanced. In target localization, a semi-global stereo matching (SGM) method is used to complete stereo matching, obtaining the disparity of the target region (ROI). A depth point cloud map is obtained through the disparity, and line fitting is performed on the depth point cloud map to locate the target conductor. This achieves the generation of a depth map from an RGB image, thereby completing the localization of the target for the live-line working robot and greatly expanding its practicality.

[0067] Figure 2 This is a system structure diagram of a target positioning system for a live-line working robot provided in one embodiment of the present invention. Figure 2 As shown, this invention provides a target localization system for a live-line working robot. The system includes: a data acquisition unit for acquiring image information containing the target and preprocessing the image information to obtain a preprocessed image; a processing unit for: using the preprocessed image as input to a pre-optimized VGG-UNet model, identifying the target, and fitting the centerline of the identified target; obtaining a disparity map based on a semi-global stereo matching method and the fitted centerline; and a localization unit for solving for the depth and position of the target based on the disparity map.

[0068] The present invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described method for locating the target of a live-line working robot.

[0069] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0070] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.

[0071] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.

Claims

1. A method for locating a work target for a live-line working robot, characterized in that, The method includes: Image information containing the task target is acquired, and the image information is preprocessed to obtain a preprocessed image; wherein, The preprocessing of the image information includes: image correction processing and image enhancement processing; Image correction includes: obtaining camera calibration parameters and distortion parameters; wherein the camera is a stereo camera; the camera calibration parameters include preset internal parameters and preset external parameters of the camera; calculating the image projection matrix based on the camera calibration parameters and distortion parameters; correcting non-coplanar line aligned stereo images to coplanar line aligned images, and performing a one-dimensional search on the epipolar line to match corresponding points of the two images, so that the row pixels of the stereo images are aligned one by one; Image enhancement processing includes: extracting an image enhancement curve from the image after image correction, which is represented as: in, Pixel enhancement results; The preset image enhancement coefficient; pixel coordinate system The light intensity value at each pixel coordinate point is obtained by performing multiple rounds of mapping on the image enhancement curve to obtain the light intensity value after each round of mapping; wherein... The formula for converting the light intensity value after each round of mapping is: in, This represents the result of the nth pixel enhancement mapping. The image enhancement coefficients are used for the nth mapping. This represents the light intensity value after the (n-1)th mapping. A pre-introduced image enhancement parameter matrix of the same size as the input image is used; the optimal image enhancement parameter matrix is ​​found based on the Zero-DCE network, and the enhanced light intensity value is obtained based on the optimal image enhancement parameter matrix; the corresponding enhanced image is obtained based on the enhanced light intensity value. The preprocessed image is used as the input image of the pre-optimized VGG-UNet model to perform task target recognition and fit the center line of the recognized task target. Disparity maps are obtained based on a semi-global stereo matching method and a fitted centerline. The depth and location of the target are obtained based on the disparity map.

2. The method according to claim 1, characterized in that, The method further includes: Constructing a pre-optimized VGG-UNet model includes: The VGG16 network and the UNet network were constructed separately, and the two models were combined to obtain the VGG-UNet network; the combination of the two models included: Remove the fully connected layers in the VGG16 network and replace them with the enhanced feature extraction network of the UNet network; Historical image information containing the task target is collected, and a VGG-UNet model for image segmentation of the task target is trained in the VGG-UNet network based on the historical image information.

3. The method according to claim 2, characterized in that, The historical image information containing the task target includes: Historical image information of the target under different sunlight angles, different weather conditions, different shooting angles, and different shooting postures; The VGG-UNet model for image segmentation of the task target, trained in the VGG-UNet network based on the historical image information, includes: The historical image information is subjected to image enhancement operations, and then the enhanced images are labeled with the target of the task to obtain a labeled dataset; The labeled dataset is imported into the VGG-UNet network for training to obtain the VGG-UNet model.

4. The method according to claim 1, characterized in that, The step of using the preprocessed image as input to the pre-optimized VGG-UNet model for target recognition and fitting the centerline of the recognized target includes: The preprocessed image is imported into the VGG-UNet model for target segmentation, and the segmented target regions are highlighted. By fitting the centerline of the highlighted area using the image's central moments, the centerline of the target in the image can be obtained.

5. The method according to claim 1, characterized in that, The process of obtaining the depth and location of the target based on the disparity map includes: Obtain the center-to-center distance, focal length, and parallax of corresponding points on the left and right sides of the two cameras; The depth value of each pixel is calculated based on the center distance between the two cameras, the focal length, and the parallax of the corresponding left and right points. Calculate the average depth value of each pixel of the target and use this average value as the distance from the target to the camera; determine the position of the target based on this distance.

6. A target positioning system for a live-line working robot, characterized in that, The system is used to execute the target localization method for a live-line working robot according to any one of claims 1-5, and the system comprises: The acquisition unit is used to acquire image information containing the task target and preprocess the image information to obtain a preprocessed image; Processing unit, used for: The preprocessed image is used as the input image of the pre-optimized VGG-UNet model to perform task target recognition and fit the center line of the recognized task target. Disparity maps are obtained based on a semi-global stereo matching method and a fitted centerline. The positioning unit is used to solve for the depth and position of the target based on the disparity map.

7. A computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the target localization method for a live-line working robot as described in any one of claims 1-5.

Images