An unmanned aerial vehicle (UAV) line following method and system based on image recognition algorithm

Abstract

The application discloses a kind of unmanned plane line simulation method and system based on image recognition algorithm, the method includes: constructing light weight transmission line wire detection model, real-time image is collected by the camera carried by unmanned plane;Light weight transmission line wire detection model identifies wire in real-time image, when the identification result is that there is wire in real-time image, adjust the focal length of camera lens;Wire image is obtained by using the camera after focal length adjustment to wire image acquisition, based on the improved Kalman filtering algorithm to wire image is handled to track wire;The distance between unmanned plane and wire is measured by the millimeter wave radar carried on unmanned plane, the flight condition of unmanned plane is controlled by taking initiative control strategy to make the distance between unmanned plane and wire keep as fixed preset distance.The application reduces the cost of unmanned plane line simulation inspection equipment, and can realize the line simulation inspection of small multi-rotor unmanned plane.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) technology, specifically a method and system for UAV line following based on image recognition algorithms. Background Technology

[0002] Transmission lines play a crucial role in power transmission within the power system. Manually operated drones struggle to inspect for defects in conductors and ground wires. Drone-based line tracking inspection involves drones autonomously following the line at a fixed distance, reducing the workload of manual drone inspections. However, accurate identification of conductors and ground wires is essential for drone-based line tracking inspection to clearly reveal defects in the photographs taken.

[0003] Current UAV line-tracking inspection methods have the following shortcomings: image recognition algorithms struggle to accurately detect small targets like wires, and because wires easily blend into the background, precise target tracking is impossible. The commonly used lidar ranging method for line tracking is costly and cannot be mounted on small multi-rotor UAVs. Although technologies utilizing UAV image recognition for wire identification and tracking exist, current UAV line-tracking inspection control algorithms still cannot accurately control flight attitude based on target detection to ensure safe flight. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a UAV line-following method based on image recognition algorithms, comprising:

[0005] A lightweight transmission line conductor detection model was constructed, and real-time images were collected using a camera mounted on a drone.

[0006] A lightweight transmission line conductor detection model identifies conductors in real-time images. When the identification result indicates that there are conductors in the real-time image, the camera lens focal length is adjusted.

[0007] The image of the conductor is acquired by using a camera with adjusted focal length. The image of the conductor is then processed based on an improved Kalman filter algorithm to track the conductor.

[0008] The distance between the drone and the wire is measured by a millimeter-wave radar mounted on the drone. An active control strategy is adopted to control the drone's flight to keep the distance between the drone and the wire at a fixed preset distance.

[0009] Furthermore, the process of constructing a lightweight transmission line conductor inspection model is as follows:

[0010] A transmission line conductor detection model is constructed. The backbone network of the transmission line conductor detection model is pruned by weight partitioning to obtain a pruned transmission line conductor detection model. The activation values ​​in the pruned transmission line conductor detection model are quantized using a two-step optimization algorithm and a clustering sparsity algorithm. Based on the quantization results of the activation values, the weights in the pruned transmission line conductor detection model are sparsely quantized to obtain a lightweight transmission line conductor detection model. The lightweight transmission line conductor detection model is then deployed on the edge computing terminal of a UAV.

[0011] Furthermore, the specific process of constructing the transmission line conductor detection model is as follows:

[0012] Obtain the image dataset of the conductor, preprocess the sample images in the image dataset, and classify and label the preprocessed sample images;

[0013] A transmission line conductor detection model is constructed by training a transfer-improved YOLOX model on a dataset of labeled images.

[0014] Furthermore, the improved YOLOX model was trained using the labeled image dataset to obtain a transmission line conductor detection model. During the training process, the Harris Eagle optimization algorithm was used to optimize the model parameters of the improved YOLOX model.

[0015] The optimization process is as follows:

[0016] Initialize the model parameters of the transfer-improved YOLOX model as an initial vector within a preset global scope;

[0017] The fitness function calculates the probability value corresponding to the target position of the initial vector. If the probability value is greater than or equal to the preset value, the target position is updated arbitrarily in the entire domain, and the fitness function calculates the probability value corresponding to the updated target position. If the probability value is less than the preset value, the optimal position of the target position is determined by the preset behavior and random walk measurement according to the magnitude of the probability value. The fitness function adopts a loss function based on the global vector.

[0018] Furthermore, the preprocessing includes cropping, rotation, contrast stretching, image enhancement, and filtering.

[0019] The contrast stretching process is represented as follows:

[0020] ;

[0021] in, and These are the grayscale values ​​of the sample image before and after stretching, respectively. and These represent the grayscale values ​​at the 2% and 97% percentiles of the grayscale distribution in the sample image before stretching;

[0022] The image enhancement process is represented as:

[0023] ;

[0024] in, It is a constant and is set according to the brightness of the sample image; For the sample image before image enhancement, the pixel values ​​are... Value represents, Adjust parameters for each pixel. For the sample image, the pixels after image enhancement. Represents the pixel adjustment function;

[0025] The filtering process is represented as follows:

[0026] ;

[0027] in, Let be the pixel value of the i-th sample image after filtering. Let be the pixel value of the i-th sample image as input. For any pixel value in the sample image before filtering, Represents the index of any position in the sample image; The set of pixels in the sample image before filtering. Let be the standard deviation of the Gaussian function. Let i be the weight vector of the i-th sample image. Represents an exponential function. This indicates taking the absolute value.

[0028] Furthermore, the specific construction method of the transfer-improved YOLOX model is as follows:

[0029] Obtain the original YOLOX model;

[0030] Some convolutional modules were built on the backbone of the original YOLOX model to replace the original four ordinary convolutional modules.

[0031] Some convolutional modules have a single structure. Partial convolutional layers and 2 It consists of ordinary convolutional layers connected in series;

[0032] A Nix module is introduced after the two C2f modules that are first in the original YOLOX backbone network. The Nix module includes a self-attention layer, a convolutional combination layer, and a feature aggregation layer.

[0033] The input features are obtained through a convolutional combination layer in the Nix module, and self-attention features are obtained through a self-attention layer in the Nix module. Finally, the convolutional and self-attention features are aggregated through a feature aggregation layer in the Nix module to obtain aggregated self-attention features. The process of obtaining aggregated self-attention features is represented as follows:

[0034] ;

[0035] in, This is a feature aggregation for self-attention features. and All are preset weight matrices. For convolutional features, This is a self-attention feature.

[0036] Furthermore, a loss function based on global vectors. Represented as:

[0037] ;

[0038] ;

[0039] ;

[0040] ;

[0041] in, Attention coefficient As a penalty amount, and All of these are hyperparameters. To improve the outlier rate of anchor boxes in the YOLOX model, For the predicted bounding box of the conductor in the sample image, These are the x and y coordinates of the prediction frame for the conductor, respectively. The true bounding box of the wire in the sample image. These are the x and y coordinates of the true bounding box of the wire in the sample image, respectively; and Let represent the width and height of the minimum bounding box of the predicted wire in the sample image, respectively. To compensate for the penalty function, The penalty function for the transfer improvement YOLOX model, This represents an exponential function.

[0042] Furthermore, the process of pruning the backbone network in the transmission line conductor detection model through weight partitioning is as follows:

[0043] ;

[0044] in, For the model output function, For the model's learnable parameters, The pixel values ​​of the input sample image. Indicates the weighted score;

[0045] Based on the weight scores, connection weights below a threshold in the backbone network are removed.

[0046] Furthermore, the process of adjusting the camera lens focal length is as follows:

[0047] Construct adjustment vector , represented as:

[0048] ;

[0049] in, These are the focal length adjustment vector, angle adjustment vector, and position adjustment vector, respectively.

[0050] The lens is adjusted based on the adjustment vector, and the target image is obtained by shooting the guide wire with the adjusted lens. The target image is evaluated to obtain the image evaluation vector, and it is determined whether the image evaluation vector reaches the set score.

[0051] When the image evaluation vector does not reach the set score, the adjustment vector is adjusted, the lens focal length is adjusted using the adjusted adjustment vector, and the target image is captured again using the adjusted lens until the image evaluation vector of the target image reaches the set score. The target image with the image evaluation vector reaching the set score is then used as the guide image.

[0052] The improved Kalman filter method for tracking wires is expressed as follows:

[0053] ;

[0054] in, This represents the state of conductor m at time t. Indicates that conductor m is in the first position. The motion function at time 1 Indicates the first The noise sequence at each time step, This represents the position of conductor m at time t. Indicates that conductor m is in the first position. Position change function at each moment Indicates the first Measurement noise at each moment;

[0055] Active control strategy is represented as:

[0056] ;

[0057] ;

[0058] in, For the drone in the next discrete moment state, For the drone at the current discrete moment state, For the drone at the current discrete moment The control input, For the drone at the current discrete moment The output status, Here is the state transition matrix. To control the input matrix, For the output matrix, This is a direct transmission matrix.

[0059] A drone line-following system based on image recognition algorithms, used in the aforementioned drone line-following method based on image recognition algorithms, includes:

[0060] The data acquisition and preprocessing module is used to acquire image datasets of overhead ground wires and conductors, and to preprocess and classify the image datasets to construct sample sets, validation sets, and test sets.

[0061] The model training and optimization module is used to load the YOLOX model and train it based on the YOLOX model, optimize the YOLOX model parameters, and obtain the transmission line conductor detection model.

[0062] The model lightweighting and deployment module is used to lightweight the transmission line conductor detection model and deploy it on the UAV computing terminal.

[0063] The drone control module is used to adjust the lens focal length to a suitable position and angle for taking pictures of the guide wire, and to control the drone to maintain a fixed distance from the guide wire for autonomous flight.

[0064] The beneficial effects of this invention are as follows:

[0065] 1. This invention introduces filtering and an attention-based Nix module into the image processing process, optimizes the model parameters using the Harris Eagle algorithm, and further reduces the model weight through weight partitioning. Ultimately, it improves the target recognition ability of the power transmission line conductor detection model for conductors with less computing power, avoids the confusion between conductors and sky in the image, and enables UAVs to follow the conductors more accurately.

[0066] 2. This invention improves the image quality captured by UAV line-following inspection by combining an improved Kalman filter algorithm with an active control strategy. Compared with existing line-following technologies, it is superior in terms of focus quality, image size, and angle, resulting in better inspection performance. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating the steps of a UAV line-following method based on an image recognition algorithm according to the present invention. Detailed Implementation

[0068] Reference Figure 1 A method for drone line following based on image recognition algorithms, comprising:

[0069] A lightweight transmission line conductor detection model was constructed, and real-time images were collected using a camera mounted on a drone.

[0070] A lightweight transmission line conductor detection model identifies conductors in real-time images. When the identification result indicates that there are conductors in the real-time image, the camera lens focal length is adjusted.

[0071] The image of the conductor is acquired by using a camera with adjusted focal length. The image of the conductor is then processed based on an improved Kalman filter algorithm to track the conductor.

[0072] The distance between the drone and the wire is measured by a millimeter-wave radar mounted on the drone. An active control strategy is adopted to control the drone's flight to keep the distance between the drone and the wire at a fixed preset distance.

[0073] Furthermore, the process of constructing a lightweight transmission line conductor inspection model is as follows:

[0074] A transmission line conductor detection model is constructed. The backbone network of the transmission line conductor detection model is pruned by weight partitioning to obtain a pruned transmission line conductor detection model. The activation values ​​in the pruned transmission line conductor detection model are quantized using a two-step optimization algorithm and a clustering sparsity algorithm. Based on the quantization results of the activation values, the weights in the pruned transmission line conductor detection model are sparsely quantized to obtain a lightweight transmission line conductor detection model. The lightweight transmission line conductor detection model is then deployed on the edge computing terminal of a UAV.

[0075] Furthermore, the specific process of constructing the transmission line conductor detection model is as follows:

[0076] Obtain the image dataset of the conductor, preprocess the sample images in the image dataset, and classify and label the preprocessed sample images;

[0077] A transmission line conductor detection model is constructed by training a transfer-improved YOLOX model on a dataset of labeled images.

[0078] Furthermore, the improved YOLOX model was trained using the labeled image dataset to obtain a transmission line conductor detection model. During the training process, the Harris Eagle optimization algorithm was used to optimize the model parameters of the improved YOLOX model.

[0079] The optimization process is as follows:

[0080] Initialize the model parameters of the transfer-improved YOLOX model as an initial vector within a preset global scope;

[0081] The fitness function calculates the probability value corresponding to the target position of the initial vector. If the probability value is greater than or equal to the preset value, the target position is updated arbitrarily in the entire domain, and the fitness function calculates the probability value corresponding to the updated target position. If the probability value is less than the preset value, the optimal position of the target position is determined by the preset behavior and random walk measurement according to the magnitude of the probability value. The fitness function adopts a loss function based on the global vector.

[0082] Furthermore, the preprocessing includes cropping, rotation, contrast stretching, image enhancement, and filtering.

[0083] The contrast stretching process is represented as follows:

[0084] ;

[0085] in, and These are the grayscale values ​​of the sample image before and after stretching, respectively. and These represent the grayscale values ​​at the 2% and 97% percentiles of the grayscale distribution in the sample image before stretching;

[0086] The image enhancement process is represented as:

[0087] ;

[0088] in, It is a constant and is set according to the brightness of the sample image; For the sample image before image enhancement, the pixel values ​​are... Value represents, Adjust parameters for each pixel. For the sample image, the pixels after image enhancement. Represents the pixel adjustment function;

[0089] The filtering process is represented as follows:

[0090] ;

[0091] in, Let be the pixel value of the i-th sample image after filtering. Let be the pixel value of the i-th sample image as input. For any pixel value in the sample image before filtering, Represents the index of any position in the sample image; The set of pixels in the sample image before filtering. Let be the standard deviation of the Gaussian function. Let i be the weight vector of the i-th sample image. Represents an exponential function. This indicates taking the absolute value.

[0092] Furthermore, the specific construction method of the transfer-improved YOLOX model is as follows:

[0093] Obtain the original YOLOX model;

[0094] Some convolutional modules were built on the backbone of the original YOLOX model to replace the original four ordinary convolutional modules.

[0095] Some convolutional modules have a single structure. Partial convolutional layers and 2 It consists of ordinary convolutional layers connected in series;

[0096] A Nix module is introduced after the two C2f modules that are first in the original YOLOX backbone network. The Nix module includes a self-attention layer, a convolutional combination layer, and a feature aggregation layer.

[0097] The input features are obtained through a convolutional combination layer in the Nix module, and self-attention features are obtained through a self-attention layer in the Nix module. Finally, the convolutional and self-attention features are aggregated through a feature aggregation layer in the Nix module to obtain aggregated self-attention features. The process of obtaining aggregated self-attention features is represented as follows:

[0098] ;

[0099] in, This is a feature aggregation for self-attention features. and All are preset weight matrices. For convolutional features, This is a self-attention feature.

[0100] Furthermore, a loss function based on global vectors. Represented as:

[0101] ;

[0102] ;

[0103] ;

[0104] ;

[0105] in, Attention coefficient As a penalty amount, and All of these are hyperparameters. To improve the outlier rate of anchor boxes in the YOLOX model, For the predicted bounding box of the conductor in the sample image, These are the x and y coordinates of the prediction frame for the conductor, respectively. The true bounding box of the wire in the sample image. These are the x and y coordinates of the true bounding box of the wire in the sample image, respectively; and Let represent the width and height of the minimum bounding box of the predicted wire in the sample image, respectively. To compensate for the penalty function, The penalty function for the transfer improvement YOLOX model, This represents an exponential function.

[0106] Furthermore, the process of pruning the backbone network in the transmission line conductor detection model through weight partitioning is as follows:

[0107] ;

[0108] in, For the model output function, For the model's learnable parameters, The pixel values ​​of the input sample image. Indicates the weighted score;

[0109] Based on the weight scores, connection weights below a threshold in the backbone network are removed.

[0110] Furthermore, the process of adjusting the camera lens focal length is as follows:

[0111] Construct adjustment vector , represented as:

[0112] ;

[0113] in, These are the focal length adjustment vector, angle adjustment vector, and position adjustment vector, respectively.

[0114] The lens is adjusted based on the adjustment vector, and the target image is obtained by shooting the guide wire with the adjusted lens. The target image is evaluated to obtain the image evaluation vector, and it is determined whether the image evaluation vector reaches the set score.

[0115] When the image evaluation vector does not reach the set score, the adjustment vector is adjusted, the lens focal length is adjusted using the adjusted adjustment vector, and the target image is captured again using the adjusted lens until the image evaluation vector of the target image reaches the set score. The target image with the image evaluation vector reaching the set score is then used as the guide image.

[0116] The improved Kalman filter method for tracking wires is expressed as follows:

[0117] ;

[0118] in, This represents the state of conductor m at time t. Indicates that conductor m is in the first position. The motion function at time 1 Indicates the first The noise sequence at each time step, This represents the position of conductor m at time t. Indicates that conductor m is in the first position. Position change function at each moment Indicates the first Measurement noise at each moment;

[0119] Active control strategy is represented as:

[0120] ;

[0121] ;

[0122] in, For the drone in the next discrete moment state, For the drone at the current discrete moment state, For the drone at the current discrete moment The control input, For the drone at the current discrete moment The output status, Here is the state transition matrix. To control the input matrix, For the output matrix, This is a direct transmission matrix.

[0123] A drone line-following system based on image recognition algorithms, used in the aforementioned drone line-following method based on image recognition algorithms, includes:

[0124] The data acquisition and preprocessing module is used to acquire image datasets of overhead ground wires and conductors, and to preprocess and classify the image datasets to construct sample sets, validation sets, and test sets.

[0125] The model training and optimization module is used to load the YOLOX model and train it based on the YOLOX model, optimize the YOLOX model parameters, and obtain the transmission line conductor detection model.

[0126] The model lightweighting and deployment module is used to lightweight the transmission line conductor detection model and deploy it on the UAV computing terminal.

[0127] The drone control module is used to adjust the lens focal length to a suitable position and angle for taking pictures of the guide wire, and to control the drone to maintain a fixed distance from the guide wire for autonomous flight.

[0128] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0129] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for UAV line-following based on image recognition algorithms, characterized in that, include: A lightweight transmission line conductor detection model was constructed, and real-time images were collected using a camera mounted on a drone. A lightweight transmission line conductor detection model identifies conductors in real-time images. When the identification result indicates that there are conductors in the real-time image, the camera lens focal length is adjusted. The image of the conductor is acquired by using a camera with adjusted focal length. The image of the conductor is then processed based on an improved Kalman filter algorithm to track the conductor. The distance between the drone and the wire is measured by a millimeter-wave radar mounted on the drone. An active control strategy is adopted to control the flight of the drone so that the distance between the drone and the wire is kept at a fixed preset distance. The process of constructing a lightweight transmission line conductor inspection model is as follows: A transmission line conductor detection model is constructed. The backbone network of the transmission line conductor detection model is pruned by weight partitioning to obtain a pruned transmission line conductor detection model. The activation values ​​in the pruned transmission line conductor detection model are quantized using a two-step optimization algorithm and a clustering sparsity algorithm. Based on the quantization results of the activation values, the weights in the pruned transmission line conductor detection model are sparsely quantized to obtain a lightweight transmission line conductor detection model. The lightweight transmission line conductor detection model is then deployed on the edge computing terminal of a UAV. The specific process of constructing a transmission line conductor detection model is as follows: Obtain the image dataset of the conductor, preprocess the sample images in the image dataset, and classify and label the preprocessed sample images; A transmission line conductor detection model is constructed by training the transfer-improved YOLOX model on the image dataset after classification and labeling. The process of adjusting the camera lens focal length is as follows: Construct adjustment vector , represented as: ； in, These are the focal length adjustment vector, angle adjustment vector, and position adjustment vector, respectively. The lens is adjusted based on the adjustment vector, and the target image is obtained by shooting the guide wire with the adjusted lens. The target image is evaluated to obtain the image evaluation vector, and it is determined whether the image evaluation vector reaches the set score. When the image evaluation vector does not reach the set score, the adjustment vector is adjusted, the lens focal length is adjusted using the adjusted adjustment vector, and the target image is captured again using the adjusted lens until the image evaluation vector of the target image reaches the set score. The target image with the image evaluation vector reaching the set score is then used as the guide image. The improved Kalman filter method for tracking wires is expressed as follows: ； in, This represents the state of conductor m at time t. Indicates that conductor m is in the first position. The motion function at time 1 Indicates the first The noise sequence at each time step, This represents the position of conductor m at time t. Indicates that conductor m is in the first position. Position change function at each moment Indicates the first Measurement noise at each moment; Active control strategy is represented as: ； ； in, For the drone in the next discrete moment state, For the drone at the current discrete moment state, For the drone at the current discrete moment The control input, For the drone at the current discrete moment The output status, Here is the state transition matrix. To control the input matrix, For the output matrix, This is a direct transmission matrix.

2. The method for UAV line following based on image recognition algorithm according to claim 1, characterized in that, The improved YOLOX model was trained using the labeled image dataset to obtain a transmission line conductor detection model. During the training process, the Harris Eagle optimization algorithm was used to optimize the model parameters of the improved YOLOX model. The optimization process is as follows: Initialize the model parameters of the transfer-improved YOLOX model as an initial vector within a preset global scope; The fitness function calculates the probability value corresponding to the target position of the initial vector. If the probability value is greater than or equal to the preset value, the target position is updated arbitrarily in the entire domain, and the fitness function calculates the probability value corresponding to the updated target position. If the probability value is less than the preset value, the optimal position of the target position is determined by the preset behavior and random walk measurement according to the magnitude of the probability value. The fitness function adopts a loss function based on the global vector.

3. A method for UAV line following based on image recognition algorithm according to claim 1, characterized in that, The preprocessing includes cropping, rotation, contrast stretching, image enhancement, and filtering. The contrast stretching process is represented as follows: ； in, and These are the grayscale values ​​of the sample image before and after stretching, respectively. and These represent the grayscale values ​​at the 2% and 97% percentiles of the grayscale distribution in the sample image before stretching; The image enhancement process is represented as: ； in, It is a constant and is set according to the brightness of the sample image; For the sample image before image enhancement, the pixel values ​​are... Value represents, Adjust parameters for each pixel. For the sample image, the pixels after image enhancement. Represents the pixel adjustment function; The filtering process is represented as follows: ； in, Let be the pixel value of the i-th sample image after filtering. Let be the pixel value of the i-th sample image as input. For any pixel value in the sample image before filtering, Represents the index of any position in the sample image; The set of pixels in the sample image before filtering. Let be the standard deviation of the Gaussian function. Let i be the weight vector of the i-th sample image. Represents an exponential function. This indicates taking the absolute value.

4. A method for UAV line following based on image recognition algorithm according to claim 2, characterized in that, The specific construction method of the transfer-improved YOLOX model is as follows: Obtain the original YOLOX model; Some convolutional modules were built on the backbone of the original YOLOX model to replace the original four ordinary convolutional modules. Some convolutional modules have a single structure. Partial convolutional layers and 2 It consists of ordinary convolutional layers connected in series; A Nix module is introduced after the two C2f modules that are first in the original YOLOX backbone network. The Nix module includes a self-attention layer, a convolutional combination layer, and a feature aggregation layer. The input features are obtained through a convolutional combination layer in the Nix module, and self-attention features are obtained through a self-attention layer in the Nix module. Finally, the convolutional and self-attention features are aggregated through a feature aggregation layer in the Nix module to obtain aggregated self-attention features. The process of obtaining aggregated self-attention features is represented as follows: ； in, This is a feature aggregation for self-attention features. and All are preset weight matrices. For convolutional features, This is a self-attention feature.

5. A method for UAV line following based on image recognition algorithm according to claim 2, characterized in that, Loss function based on global vector Represented as: ； ； ； ； in, Attention coefficient As a penalty amount, and All of these are hyperparameters. To improve the outlier rate of anchor boxes in the YOLOX model, For the predicted bounding box of the conductor in the sample image, These are the x and y coordinates of the prediction frame for the conductor, respectively. The true bounding box of the wire in the sample image. These are the x and y coordinates of the true bounding box of the wire in the sample image, respectively; and Let represent the width and height of the minimum bounding box of the predicted wire in the sample image, respectively. To compensate for the penalty function, The penalty function for the transfer improvement YOLOX model, This represents an exponential function.

6. A method for UAV line following based on image recognition algorithm according to claim 1, characterized in that, The specific process of pruning the backbone network in the transmission line conductor detection model through weight partitioning is as follows: ； in, For the model output function, For the model's learnable parameters, The pixel values ​​of the input sample image. Indicates the weighted score; Based on the weight scores, connection weights below a threshold in the backbone network are removed.

7. A UAV line-following system based on image recognition algorithm, used to implement the UAV line-following method based on image recognition algorithm as described in any one of claims 1 to 6, characterized in that, include: The data acquisition and preprocessing module is used to acquire image datasets of overhead ground wires and conductors, and to preprocess and classify the image datasets to construct sample sets, validation sets, and test sets. The model training and optimization module is used to load the YOLOX model and train it based on the YOLOX model, optimize the YOLOX model parameters, and obtain the transmission line conductor detection model. The model lightweighting and deployment module is used to lightweight the transmission line conductor detection model and deploy it on the UAV computing terminal. The drone control module is used to adjust the lens focal length to a suitable position and angle for taking pictures of the guide wire, and to control the drone to maintain a fixed distance from the guide wire for autonomous flight.

Images