Unmanned aerial vehicle autonomous inspection method based on AI technology

Abstract

The invention relates to an unmanned aerial vehicle autonomous inspection method based on an AI technology, and the method comprises the following steps: S1, capturing a high-resolution image through an airborne camera in an inspection process, carrying out the spatial calibration of the image through IMU and LiDAR data, and guaranteeing the precise correspondence between the captured image data and a physical environment; s2, extracting texture features in the image by using a Gabor filter, acquiring edge features in the image by using a Canny edge detection algorithm, and detecting a linear structure by using Hough transform to assist in identifying a lead; s3, an improved DeepLab model is adopted to segment a lead area; s4, performing real-time tracking on the identified lead position by using a Kalman filter, and introducing a dynamic Bayesian network DBN to perform short-term change prediction on the lead position; and S5, generating a map of the current environment based on SLAM, and optimizing the inspection path of the unmanned aerial vehicle based on an RRT improved algorithm. Accurate wire inspection is achieved, it is ensured that operation is conducted along a safe route of a wire, efficiency is improved, and the risk of inspection is reduced.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle inspection, and in particular to an autonomous unmanned aerial vehicle inspection method based on AI technology. Background Art

[0002] In modern power transmission systems, power line inspection is a key link in maintaining the stability and reliability of the power grid. Traditional manual inspection methods are inefficient, costly, and accompanied by certain safety risks. In recent years, with the development of unmanned driving technology and artificial intelligence, drones have gradually become one of the effective tools in the field of power inspection. However, power lines usually pass through various complex and changeable environments, including mountainous areas, forests, and urban areas, which poses severe challenges to the autonomous flight and obstacle avoidance capabilities of drones. Summary of the invention

[0003] In order to solve the above problems, the purpose of the present invention is to provide an autonomous inspection method of a drone based on AI technology to achieve accurate wire inspection, ensure operations along a safe route of the wire, improve efficiency and reduce inspection risks.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] An autonomous inspection method for drones based on AI technology includes the following steps:

[0006] S1: During the inspection process, high-resolution images are captured using an onboard camera, and spatial calibration of the images is performed using IMU and LiDAR data to ensure that the captured image data accurately corresponds to the physical environment;

[0007] S2: Use Gabor filter to extract texture features in the image, enhance the saliency of the wire in the image, apply Canny edge detection algorithm to obtain edge features in the image, and use Hough transform to detect straight line structure to assist in identifying the wire;

[0008] S3: Use the improved DeepLab model to segment the wire area;

[0009] S4: Use Kalman filter to track the identified wire position in real time, and introduce dynamic Bayesian network DBN to predict the short-term change of the wire position;

[0010] S5: Generate a map of the current environment based on SLAM to update the acquired path information in real time, and optimize the inspection path of the drone based on the RRT improved algorithm to ensure the smoothness and safety of the path while avoiding obstacles.

[0011] Furthermore, S1 is specifically:

[0012] First, obtain the real-time attitude information of the drone based on the IMU data, and use quaternions to represent and calculate the attitude:

[0013]

[0014] Among them, q t is the current quaternion, w t is the angular velocity vector, Δ t is the time interval, Represents the cross product operation;

[0015] Based on the attitude information obtained by the IMU, the LiDAR point cloud data is converted from the drone coordinate system to the world coordinate system:

[0016] p world =R·p drone +t;

[0017] Among them, p drone is the point coordinate in the drone coordinate system, R is the rotation matrix; t is the translation vector; p world is the point coordinate in the world coordinate system;

[0018] Project the point cloud data onto the image plane, set the camera intrinsic parameter matrix to K, and the perspective projection formula is:

[0019]

[0020] Combine the depth data of LiDAR with the RGB image of the camera, and use the projection coordinates u to complete the matching of image and depth information.

[0021] Furthermore, S2 is specifically:

[0022] A set of Gabor filters with different directions and scales are defined; these Gabor filters are applied to the image to obtain an enhanced texture image;

[0023]

[0024] Where λ is the wavelength, θ is the direction of the Gaussian window, which controls the directivity of the filter; ψ is the phase shift; σ is the standard deviation of the Gaussian envelope; γ is the spatial aspect ratio; (x, y) are the image coordinates;

[0025] The Gabor filter output is used as the input of feature enhancement for subsequent edge detection;

[0026] Use a Gaussian filter to smooth the image to remove noise, calculate the image gradient magnitude and direction; suppress non-maxima along the gradient direction, and apply high and low double thresholds to identify and connect edges;

[0027] Hough transform is used to identify straight line structures in images, which assists in the identification and real-time tracking of wires.

[0028] Furthermore, the improved DeepLab model includes an encoder part and a decoder part, as follows:

[0029] The encoder part uses a pre-trained deep convolutional neural network as the backbone network to extract features and convert the input image into a feature map;

[0030] Apply dilated convolutions in feature maps to expand the receptive field without increasing computational cost:

[0031]

[0032] Among them, y(i) is the value of the output feature map at position i; S is the convolution kernel; s is the offset relative to the center of the kernel; r is the hole rate; I is the input feature map; k(s) is the weight matrix;

[0033] We use the dilated spatial pyramid pooling (ASPP) to apply dilated convolutions to different dilation rates to capture multi-scale contexts. ASPP combines multi-scale features and fuses the features of convolutions with multiple dilation rates, enabling the model to handle objects of different sizes.

[0034] The decoder combines high-level semantic features from the encoder with low-level spatial details to achieve higher-precision segmentation.

[0035] Extract low-level features directly from the encoder part and fuse them in the decoder through skip connections;

[0036] The encoded low-resolution feature map is progressively upsampled to the resolution of the original input image; upsampling uses bilinear interpolation to enlarge the feature map.

[0037] Furthermore, we adopt the dilated spatial pyramid pooling (ASPP) and apply dilated convolution to different dilation rates to capture multi-scale contexts, as follows:

[0038] The ASPP module is constructed by combining multi-layer dilated convolution and global average pooling, and added to the feature extraction part of the model;

[0039] The different dilation rates used in ASPP help capture multi-scale features; each different dilation rate corresponds to processing contextual information of different scales;

[0040] Perform average pooling on the entire feature map and output a single global feature value, which represents the global features of the entire image;

[0041] The ASPP output is progressively upsampled to the original image resolution through the decoder and the final semantic segmentation prediction is generated.

[0042] Furthermore, S4 is specifically:

[0043] Real-time tracking via Kalman filtering to get real-time position updates:

[0044] Define the state vector where x k ,y k is the two-dimensional position of the wire; Indicates the horizontal and vertical speeds of the wire movement;

[0045] Define the state transfer matrix A and the observation matrix H, initialize the process noise covariance matrix Q and the observation noise covariance matrix R;

[0046] Use the previous state to estimate the current state and covariance:

[0047]

[0048] Where, B is the control input matrix; is the predicted state vector, representing the prior estimate at time k; u k is the control input vector at the current time k; is the predicted error covariance matrix;

[0049] Update the state and covariance based on the current observation data;

[0050]

[0051] Among them, K k is the Kalman gain matrix; z k is the actual observation vector at the current moment k;

[0052] The current position of the wire is updated and output in real time; in the case of uncertain or sudden position changes, DBN is used to predict potential changes.

[0053] Furthermore, DBN is used to predict potential changes as follows:

[0054] Define state S t and observation O t , establish the state transition probability P(S t+1 |S t ) and the observation probability P(O t |S t );

[0055] Use historical data to train DBN parameters so that it can model the temporal variation characteristics of wire positions;

[0056] In the case of uncertain or sudden position, real-time observation data is input and a forward algorithm is executed to infer the probability of future states;

[0057] P(S t+1 |O 1：t )=∫P(S t+1 |S t )·P(S t |O 1：t )dS t

[0058] Output short-term position change prediction of the conductor to supplement the real-time tracking capability of the Kalman filter;

[0059] Update future state predictions based on current observations.

[0060] Furthermore, S5 is specifically:

[0061] S51: Real-time mapping of the drone’s environment and self-positioning, using cameras and LiDAR sensors to acquire data and ORB-SLAM to generate accurate environmental maps;

[0062] S52: Use the map generated by SLAM to detect obstacle updates in real time, avoid collisions by improving the dynamic path adjustment mechanism of the RRT algorithm, ensure that the planned path maintains a safety threshold from all known obstacles, and regenerate the local path in real time.

[0063] Further, S51 is as follows:

[0064] Use cameras to obtain visual data and LiDAR to obtain 3D point cloud data. The camera provides color and brightness information, and LiDAR provides distance and depth information.

[0065] The ORB algorithm is used to extract feature points.

[0066] Based on the extracted feature points, the pose of the drone is estimated by matching between frames and solving the P3P problem.

[0067] Generate sparse point cloud map by triangulation calculation of feature points;

[0068] Construct and optimize factor graphs to improve map accuracy and consistency;

[0069] Detect and process loops, identify the paths traveled through the BoW model, and make closed-loop adjustments to optimize the overall map shape and reduce cumulative errors.

[0070] Furthermore, the RRT algorithm is improved as follows:

[0071] Create an exploration tree from the starting point to the end point based on RRT;

[0072] Expand the tree by randomly sampling and connecting to the nearest reachable node;

[0073] Keep the path length and cost of any node in the tree to a minimum;

[0074] For each new node q new Find its nearby nodes q near ;

[0075] If reconnecting these nodes can reduce the total cost, re-plan the connection relationship in the tree;

[0076] Use B-spline curves to smooth the original path to ensure the drone flies naturally;

[0077]

[0078] Among them, N i，k (t) is the B-spline basis function, and Pi is the i-th control point.

[0079] The present invention has the following beneficial effects:

[0080] 1. While ensuring the efficient inspection capability of the UAV, the present invention significantly improves its autonomous flight and obstacle avoidance capabilities in complex environments, adapts to the changing actual field environment, and greatly guarantees the efficiency of power inspection tasks; and utilizes the posture and depth information provided by IMU and LiDAR to achieve efficient and accurate image space calibration, greatly improving the reliability and accuracy of inspections.

[0081] 2. The present invention uses Kalman filtering for real-time tracking to obtain robust real-time position updates. In the case of uncertain or sudden position changes, DBN is used to predict potential changes and improve the durability and reliability of the system. The filtering and prediction information are integrated to provide accurate decision support for future positions, so that the UAV can accurately respond to changes in the position of the wire in a complex environment, ensuring high efficiency and accuracy in real-time applications.

[0082] 3. The present invention utilizes ORB-SLAM for accurate environmental modeling and self-positioning, supplemented by improved RRT* for real-time and efficient path planning and obstacle avoidance, and is capable of performing efficient and autonomous inspection tasks in complex and unknown environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0084] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0085] refer to Figure 1 In this embodiment, an autonomous inspection method of a drone based on AI technology includes the following steps:

[0086] S1: During the inspection process, high-resolution images are captured using an onboard camera, and spatial calibration of the images is performed using IMU and LiDAR data to ensure that the captured image data accurately corresponds to the physical environment;

[0087] S2: Use Gabor filter to extract texture features in the image, enhance the saliency of the wire in the image, apply Canny edge detection algorithm to obtain edge features in the image, and use Hough transform to detect straight line structure to assist in identifying the wire;

[0088] S3: Use the improved DeepLab model to segment the wire area;

[0089] S4: Use Kalman filter to track the identified wire position in real time, and introduce dynamic Bayesian network DBN to predict the short-term change of the wire position;

[0090] S5: Generate a map of the current environment based on SLAM to update the acquired path information in real time, and optimize the inspection path of the drone based on the RRT improved algorithm to ensure the smoothness and safety of the path while avoiding obstacles.

[0091] In this embodiment, S1 is specifically:

[0092] First, obtain the real-time attitude information of the drone based on the IMU data, and use quaternions to represent and calculate the attitude:

[0093]

[0094] Among them, q t is the current quaternion, w t is the angular velocity vector, Δ t is the time interval, Represents the cross product operation;

[0095] Based on the attitude information obtained by the IMU, the LiDAR point cloud data is converted from the drone coordinate system to the world coordinate system:

[0096] p world =R·p drone +t;

[0097] Among them, p drone is the point coordinate in the drone coordinate system, R is the rotation matrix; t is the translation vector; pworld is the point coordinate in the world coordinate system;

[0098] Project the point cloud data onto the image plane, set the camera intrinsic parameter matrix to K, and the perspective projection formula is:

[0099]

[0100] Combine the depth data of LiDAR with the RGB image of the camera, and use the projection coordinate u to complete the matching of image and depth information. (The image output after calibration is "scene image + depth map", that is, each image pixel exactly matches the 3D depth information. Such calibration provides higher accuracy in the subsequent image segmentation and wire recognition process)

[0101] In this embodiment, S2 is specifically:

[0102] A set of Gabor filters with different directions and scales are defined; these Gabor filters are applied to the image to obtain an enhanced texture image;

[0103]

[0104] Where λ is the wavelength, θ is the direction of the Gaussian window, which controls the directivity of the filter; ψ is the phase shift; σ is the standard deviation of the Gaussian envelope; γ is the spatial aspect ratio; (x, y) are the image coordinates;

[0105] The Gabor filter output is used as the input of feature enhancement for subsequent edge detection;

[0106] Use a Gaussian filter to smooth the image to remove noise, calculate the image gradient magnitude and direction; suppress non-maxima along the gradient direction, and apply high and low double thresholds to identify and connect edges;

[0107] Hough transform is used to identify straight line structures in images, which assists in the identification and real-time tracking of wires.

[0108] In this embodiment, the improved DeepLab model includes an encoder part and a decoder part, as follows:

[0109] The encoder part uses a pre-trained deep convolutional neural network (usually ResNet or Xception) as the backbone network to extract features and convert the input image into a feature map;

[0110] Apply dilated convolutions in feature maps to expand the receptive field without increasing computational cost:

[0111]

[0112] Among them, y(i) is the value of the output feature map at position i; S is the convolution kernel; s is the offset relative to the center of the kernel; r is the hole rate; I is the input feature map; k(s) is the weight matrix;

[0113] We use the dilated spatial pyramid pooling (ASPP) to apply dilated convolutions to different dilation rates to capture multi-scale contexts. ASPP combines multi-scale features and fuses the features of convolutions with multiple dilation rates, enabling the model to handle objects of different sizes.

[0114] The decoder combines high-level semantic features from the encoder with low-level spatial details to achieve higher-precision segmentation.

[0115] Extract low-level features directly from the encoder part and fuse them in the decoder through skip connections;

[0116] The encoded low-resolution feature map is progressively upsampled to the resolution of the original input image; upsampling uses bilinear interpolation to enlarge the feature map.

[0117] In this embodiment, the atrous spatial pyramid pooling (ASPP) is used to apply atrous convolution to different dilation rates to capture multi-scale contexts, as follows:

[0118] The ASPP module is constructed by combining multi-layer dilated convolution and global average pooling, and added to the feature extraction part of the model;

[0119] The different dilation rates used in ASPP help capture multi-scale features; each different dilation rate corresponds to processing contextual information of different scales;

[0120] Perform average pooling on the entire feature map and output a single global feature value, which represents the global features of the entire image;

[0121] The ASPP output is progressively upsampled to the original image resolution through the decoder and the final semantic segmentation prediction is generated.

[0122] In this embodiment, S4 is specifically:

[0123] Real-time tracking via Kalman filtering to get real-time position updates:

[0124] Define the state vector where x k ,y k is the two-dimensional position of the wire; surface

[0125] Indicates the horizontal and vertical speeds of the conductor movement;

[0126] Define the state transfer matrix A and the observation matrix H, initialize the process noise covariance matrix Q and the observation noise covariance matrix R;

[0127] Use the previous state to estimate the current state and covariance:

[0128]

[0129] Where, B is the control input matrix; is the predicted state vector, representing the prior estimate at time k; u k is the control input vector at the current time k; is the predicted error covariance matrix;

[0130] Update the state and covariance based on the current observation data;

[0131]

[0132] Among them, K k is the Kalman gain matrix; z k is the actual observation vector at the current moment k;

[0133] The current position of the wire is updated and output in real time; in the case of uncertain or sudden position changes, DBN is used to predict potential changes.

[0134] In this embodiment, DBN is used to predict potential changes as follows:

[0135] Define state S t and observation O t , establish the state transition probability P(S t+1 |S t ) and the observation probability P(O t |S t );

[0136] Use historical data to train DBN parameters so that it can model the temporal variation characteristics of wire positions;

[0137] In the case of uncertain or sudden position, real-time observation data is input and a forward algorithm is executed to infer the probability of future states;

[0138] P(S t+1 |O 1：t )=∫P(S t+1 |S t )·P(S t |O 1：t )dS t

[0139] Output short-term position change prediction of the conductor to supplement the real-time tracking capability of the Kalman filter;

[0140] Update future state predictions based on current observations.

[0141] In this embodiment, S5 is specifically:

[0142] S51: Real-time mapping of the drone’s environment and self-positioning, using cameras and LiDAR sensors to acquire data and ORB-SLAM to generate accurate environmental maps;

[0143] S52: Use the map generated by SLAM to detect obstacles in real time and update them. Improve the dynamic path adjustment mechanism of the RRT algorithm to avoid collisions, ensure that the planned path maintains a safety threshold from all known obstacles, and regenerate the local path in real time.

[0144] In this embodiment, S51 is specifically as follows:

[0145] Use cameras to obtain visual data and LiDAR to obtain 3D point cloud data. The camera provides color and brightness information, and LiDAR provides distance and depth information.

[0146] The ORB algorithm is used to extract feature points.

[0147] Based on the extracted feature points, the pose of the drone is estimated by matching between frames and solving the P3P problem.

[0148] Generate sparse point cloud map by triangulation calculation of feature points;

[0149] Construct and optimize factor graphs to improve map accuracy and consistency;

[0150] Detect and process loops, identify the paths traveled through the BoW (Bag of Words) model, and make closed-loop adjustments to optimize the overall map shape and reduce cumulative errors.

[0151] In this embodiment, the RRT algorithm is improved as follows:

[0152] Create an exploration tree from the starting point to the end point based on RRT;

[0153] Expand the tree by randomly sampling and connecting to the nearest reachable node;

[0154] Keep the path length and cost of any node in the tree to a minimum;

[0155] For each new node q new Find its nearby nodes q near ;

[0156] If reconnecting these nodes can reduce the total cost, re-plan the connection relationship in the tree;

[0157] Use B-spline curves to smooth the original path to ensure the drone flies naturally;

[0158]

[0159] Among them, N i，k (t) is the B-spline basis function, and Pi is the i-th control point.

[0160] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention 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, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0161] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0162] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0163] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0164] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. An autonomous inspection method for drones based on AI technology, characterized in that: The following steps are involved: S1: During the inspection process, high-resolution images are captured using an onboard camera, and spatial calibration of the images is performed using IMU and LiDAR data to ensure that the captured image data accurately corresponds to the physical environment; S2: Use Gabor filter to extract texture features in the image, apply Canny edge detection algorithm to obtain edge features in the image, and use Hough transform to detect straight line structure to assist in identifying wires; S3: Use the improved DeepLab model to segment the wire area; S4: Use Kalman filter to track the identified wire position in real time, and introduce dynamic Bayesian network DBN to predict the short-term change of the wire position; S5: Generate a map of the current environment based on SLAM, and optimize the inspection path of the drone based on the RRT improved algorithm to ensure the smoothness and safety of the path while avoiding obstacles.

2. The autonomous inspection method of unmanned aerial vehicle based on AI technology according to claim 1 is characterized in that: The S1 is specifically: First, obtain the real-time attitude information of the drone based on the IMU data, and use quaternions to represent and calculate the attitude: Among them, q t is the current quaternion, w t is the angular velocity vector, Δ t is the time interval, Represents the cross product operation; Based on the attitude information obtained by the IMU, the LiDAR point cloud data is converted from the drone coordinate system to the world coordinate system: p world =R·p drone +t; Among them, p drone is the point coordinate in the drone coordinate system, R is the rotation matrix; t is the translation vector; p world is the point coordinate in the world coordinate system; Project the point cloud data onto the image plane, set the camera intrinsic parameter matrix to K, and the perspective projection formula is: Combine the depth data of LiDAR with the RGB image of the camera, and use the projection coordinates u to complete the matching of image and depth information.

3. The autonomous inspection method of unmanned aerial vehicle based on AI technology according to claim 1 is characterized in that: The S2 is specifically: A set of Gabor filters with different directions and scales are defined; these Gabor filters are applied to the image to obtain an enhanced texture image; Where λ is the wavelength, θ is the direction of the Gaussian window, which controls the directivity of the filter; ψ is the phase shift; σ is the standard deviation of the Gaussian envelope; γ is the spatial aspect ratio; (x, y) are the image coordinates; The Gabor filter output is used as the input of feature enhancement for subsequent edge detection; Use a Gaussian filter to smooth the image to remove noise, calculate the image gradient magnitude and direction; suppress non-maxima along the gradient direction, and apply high and low double thresholds to identify and connect edges; Hough transform is used to identify straight line structures in images, which assists in the identification and real-time tracking of wires.

4. The autonomous inspection method of unmanned aerial vehicle based on AI technology according to claim 1 is characterized in that: The improved DeepLab model includes an encoder part and a decoder part, as follows: The encoder part uses a pre-trained deep convolutional neural network as the backbone network to extract features and convert the input image into a feature map; Apply dilated convolutions in feature maps to expand the receptive field without increasing computational cost: Among them, y(i) is the value of the output feature map at position i; S is the convolution kernel; s is the offset relative to the center of the kernel; r is the hole rate; I is the input feature map; k(s) is the weight matrix; We use the dilated spatial pyramid pooling (ASPP) to apply dilated convolutions to different dilation rates to capture multi-scale contexts. ASPP combines multi-scale features and fuses the features of convolutions with multiple dilation rates, enabling the model to handle objects of different sizes. The decoder combines high-level semantic features from the encoder with low-level spatial details to achieve higher-precision segmentation. Extract low-level features directly from the encoder part and fuse them in the decoder through skip connections; The encoded low-resolution feature map is progressively upsampled to the resolution of the original input image; upsampling uses bilinear interpolation to enlarge the feature map.

5. The method for autonomous inspection of unmanned aerial vehicles based on AI technology according to claim 4 is characterized in that: The hollow spatial pyramid pooling ASPP is used to apply hollow convolutions with different expansion rates to capture multi-scale contexts, as follows: The ASPP module is constructed by combining multi-layer dilated convolution and global average pooling, and added to the feature extraction part of the model; The different dilation rates used in ASPP help capture multi-scale features; each different dilation rate corresponds to processing contextual information of different scales; Perform average pooling on the entire feature map and output a single global feature value, which represents the global features of the entire image; The ASPP output is progressively upsampled to the original image resolution through the decoder and the final semantic segmentation prediction is generated.

6. The method for autonomous inspection of unmanned aerial vehicles based on AI technology according to claim 1 is characterized in that: The S4 is specifically: Real-time tracking via Kalman filtering to get real-time position updates: Define the state vector where x k ,y k is the two-dimensional position of the wire; Indicates the horizontal and vertical speeds of the wire movement; Define the state transfer matrix A and the observation matrix H, initialize the process noise covariance matrix Q and the observation noise covariance matrix R; Use the previous state to estimate the current state and covariance: Where, B is the control input matrix; is the predicted state vector, representing the prior estimate at time k; u k is the control input vector at the current time k; is the predicted error covariance matrix; Update the state and covariance based on the current observation data; Among them, K k is the Kalman gain matrix; z k is the actual observation vector at the current moment k; The current position of the wire is updated and output in real time; in the case of uncertain or sudden position changes, DBN is used to predict potential changes.

7. The autonomous inspection method of unmanned aerial vehicle based on AI technology according to claim 6 is characterized in that: The use of DBN to predict potential changes is as follows: Define state S t and observation O t , establish the state transition probability P(S t+1 |S t ) and the observation probability P(O t |S t ); Use historical data to train the parameters of DBN so that it can model the temporal variation characteristics of the wire position; In the case of uncertain or sudden position, real-time observation data is input and a forward algorithm is executed to infer the probability of future states; P(S t+1 |O 1：t )=∫P(S t+1 |S t )·P(S t |O 1：t )dS t Output short-term position change prediction of the conductor to supplement the real-time tracking capability of the Kalman filter; Update future state predictions based on current observations.

8. The method for autonomous inspection of unmanned aerial vehicles based on AI technology according to claim 1 is characterized in that: The S5 is specifically: S51: Real-time mapping of the drone’s environment and self-positioning, using cameras and LiDAR sensors to acquire data and ORB-SLAM to generate accurate environmental maps; S52: Use the map generated by SLAM to detect obstacles in real time and update them. Improve the dynamic path adjustment mechanism of the RRT algorithm to avoid collisions, ensure that the planned path maintains a safety threshold from all known obstacles, and regenerate the local path in real time.

9. The method for autonomous inspection of unmanned aerial vehicles based on AI technology according to claim 8 is characterized in that: The S51 is specifically as follows: Use cameras to obtain visual data and LiDAR to obtain 3D point cloud data. The camera provides color and brightness information, and LiDAR provides distance and depth information. The ORB algorithm is used to extract feature points. Based on the extracted feature points, the pose of the drone is estimated by matching between frames and solving the P3P problem. Generate sparse point cloud map by triangulation calculation of feature points; Construct and optimize factor graphs to improve map accuracy and consistency; Detect and process loops, identify the paths traveled through the BoW model, and make closed-loop adjustments to optimize the overall map shape and reduce cumulative errors.

10. The autonomous inspection method of unmanned aerial vehicle based on AI technology according to claim 8, characterized in that: The improved RRT algorithm is as follows: Create an exploration tree from the starting point to the end point based on RRT; Expand the tree by randomly sampling and connecting to the nearest reachable node; Keep the path length and cost of any node in the tree to a minimum; For each new node q new Find its nearby nodes q near ; If reconnecting these nodes can reduce the total cost, re-plan the connection relationship in the tree; Use B-spline curves to smooth the original path to ensure the drone flies naturally; Among them, N i，k (t) is the B-spline basis function, and Pi is the i-th control point.

Images