Unmanned aerial vehicle inspection control system based on binocular vision

Abstract

The invention discloses an unmanned aerial vehicle inspection control system based on binocular vision, and relates to the technical field of unmanned aerial vehicle inspection, a binocular vision acquisition module is used for acquiring a to-be-inspected target image, detecting shielding and degree, and extracting a target feature and a distribution rule when an unmanned aerial vehicle inspects according to a preset route; the information acquisition module is used for acquiring a target position, peripheral complementary target distribution, unmanned aerial vehicle pose and kinematics parameters when it is detected that the shielding degree of the target point exceeds a set value. Complete coverage of the two types of target sub-regions can be ensured, the flight trajectory curvature of the second type of target sub-regions can be minimum, and stable flight is ensured.

Description

Technical Field

[0001] This invention belongs to the field of drone inspection technology, specifically a drone inspection control system based on binocular vision. Background Technology

[0002] In fields such as power, oil and gas, construction, and security, drone inspections have become a core means of replacing manual labor in large-scale, high-risk target monitoring due to their advantages of high efficiency, flexibility, low cost, and non-contact operation.

[0003] However, pre-set route inspection faces a key bottleneck in practical applications: the occlusion problem of the targets to be inspected. Due to the complexity of the inspection scenario, some targets in the pre-set route may be obscured by obstacles, preventing the drone from directly acquiring a complete image of the target. In this case, traditional solutions often rely on post-inspection—that is, after marking the obscured targets, the route is manually replanned or the drone is operated to return for re-inspection. This approach not only increases inspection costs and time consumption but may also lose its timeliness due to the delayed re-inspection. Summary of the Invention

[0004] The purpose of this invention is to provide a binocular vision-based UAV inspection and control system, which solves the technical problem of balancing comprehensive coverage of complementary targets and ensuring flight stability in the prior art.

[0005] The drone inspection and control system based on binocular vision includes: The binocular vision acquisition module is used to acquire images of the target to be inspected when the UAV is inspecting along a preset route, detect the degree of occlusion, and extract the target features and distribution patterns. The information acquisition module is used to acquire the target location, the distribution of surrounding complementary targets, the UAV pose and kinematic parameters when the occlusion degree of the detected target point exceeds the set value. The strategy selection module, based on the acquired information, identifies the occlusion type of the target point and the visible angle range of the complementary target point, divides the detection sub-region containing the first / second type of complementary targets, and selects the corresponding inspection strategy from the preset strategy library; The path planning module is used to call the model according to the strategy, generate a path for the first type of target sub-region based on the distribution density gradient and direction; generate an initial path for the second type of target sub-region based on the height difference with the UAV, combine horizontal distance correction to minimize the trajectory curvature, and integrate to obtain the optimal path. The flight control module is used to control the UAV to inspect complementary targets along the optimal path and complete the full observation of the originally obscured targets.

[0006] Furthermore, the specific tasks of the path planning module are as follows: Based on the selected inspection strategy, this system invokes a corresponding path planning model that considers the UAV's flight trajectory, binocular vision focus adjustment, and spatial correlation of target points. For the sub-region containing the first type of complementary target points, it analyzes the spatial distribution density gradient of each target and generates a movement path within the sub-region in order from high-density areas to low-density areas, with the path direction consistent with the main distribution direction of the targets. For the sub-region containing the second type of complementary target points, it calculates the relative height difference between each target and the current position of the UAV, generates an initial path in order from smallest to largest height difference, and then smooths the path by combining the horizontal distance between targets to minimize the curvature of the flight trajectory between adjacent targets. By integrating the movement paths of each sub-region, the optimal inspection sequence and flight path of complementary target points are obtained.

[0007] Furthermore, the binocular vision acquisition module is specifically used for: The three-dimensional image information of the target under inspection is captured in real time by a binocular camera. An edge detection algorithm is used to identify the occlusion contour and calculate the proportion of the occluded area to the total area of ​​the target to determine the degree of occlusion. The three-dimensional spatial coordinates, geometric dimensions and arrangement direction features of the target are extracted as distribution pattern information. The arrangement direction features are obtained by detecting the main distribution axis of the target through Hough transform.

[0008] Furthermore, when the path planning module generates a path for the first type of target sub-region based on the distribution density gradient and orientation, it addresses local conflicts between the distribution density gradient and orientation by including: When initially generating the basic path along the main direction, mark the conflict segments in the path whose deviation distance from the local high-density cluster exceeds a set value; For conflict segments, branch paths parallel to the main direction are generated to cover high-density clusters. The branch paths are connected to the base paths through arc-shaped transition segments, and the rate of change of the turning angle of the transition segments is calculated. If the rate of change of steering angle exceeds the attitude stability threshold of the UAV, shorten the branch path length and increase the number of transition segments, recalculate the rate of change of steering angle until it meets the threshold requirement; After each round of adjustments, the deviation between the overall path and the target main direction is verified. If the deviation does not decrease after two consecutive rounds of adjustments, the local high-density cluster is divided into an independent sub-region that still belongs to the first type of target sub-region, and a separate path is planned. The verification is repeated until the conflict is resolved.

[0009] Furthermore, when the path planning module generates and corrects the initial path for the second type of target sub-regions by sorting them according to their height differences, it also includes: After the initial path is generated, the ratio of horizontal distance to height difference for each pair of adjacent targets is calculated, and abnormal target pairs with ratios exceeding a set threshold are marked. For marked abnormal target pairs, other targets within a certain range around them are retrieved as potential transition points, and branch path schemes connecting the two targets through the transition points are constructed. The energy consumption difference between this scheme and the original straight path is calculated. If the energy consumption of the branch path is lower than that of the original path, the branch path is used to replace the original connection method, the transition point is included in the path sequence, and the ratio of all adjacent targets in the new sequence is recalculated. The marking and replacement steps are repeated. After each round of replacement, it is verified whether the total inspection time of the overall path has been shortened. If the total time does not improve after two consecutive rounds of replacement, the loop is stopped, and the current optimized path is retained. Finally, the turnaround rate of the optimized path is detected. If the number of turnarounds in the same area exceeds a set value, the spatial partitioning algorithm is called to cluster the targets by horizontal distance and prioritize sorting them by height difference within the cluster to achieve a synergy between local compactness and global efficiency.

[0010] Furthermore, when integrating the two types of sub-region paths, the path planning module also includes: After initial integration, the trajectory tangent direction and the current motion vector of the UAV at the junction of the two types of paths are obtained, the direction change angle is calculated, and the transition segment where the direction change angle exceeds the set value is marked; For the marked transition segment, a virtual guide point is introduced. This guide point is located on the angle bisector of the angle between the tangents of the two types of paths, at a certain distance from the connection point, and a transition path connected by the guide point is constructed. Calculate the rate of curvature change of the transition path. If the rate of curvature change exceeds the threshold of the UAV's attitude adjustment capability, adjust the position of the guide point, increase its distance from the connection point, reconstruct the transition path, and repeat the calculation until the rate of curvature change meets the standard. The transition path is optimized by combining real-time airflow disturbance data. When the airflow disturbance intensity exceeds the set level, the number of guiding points is increased so that the transition path is a multi-segment smooth curve, with the curvature center of each segment pointing in the opposite direction of the airflow disturbance. After each round of optimization, the connection accuracy between the transition section and the two types of paths is checked, that is, the deviation value between the endpoint of the transition path and the original connection point. If the deviation value does not decrease after two consecutive optimizations, a dynamic segmentation strategy that is only for cross-type path connection scenarios is activated. The transition section is divided into different sections according to the intensity of airflow disturbance. Each section adopts an appropriate curve type. The process is repeated until the direction change angle, curvature change rate and connection accuracy all meet the requirements.

[0011] Furthermore, when the binocular vision acquisition module detects occlusion and its degree, it includes: The target stereo image pairs are acquired by a binocular camera, and the disparity map is calculated to obtain the target depth information. A pixel-level occlusion detection algorithm is adopted to count the proportion of pixels in the occluded area to the total number of pixels in the target. The proportion range is divided into three levels: mild occlusion, moderate occlusion, and severe occlusion. Mild occlusion means that the occluded area does not affect the recognition of the target's core features, while severe occlusion means that the target's core features are completely invisible.

[0012] Furthermore, the preset strategy library of the strategy selection module includes: For the dense coverage strategy for the combination of mild occlusion and the first type of complementary target, the path is set to a multi-round cross-coverage mode to improve the integrity of feature acquisition; for the efficient detour strategy for the combination of severe occlusion and the second type of complementary target, the path is based on covering the key complementary target with the shortest distance; when matching the strategy, priority is given to ensuring that the visual angle of the complementary target meets the baseline requirements of binocular visual stereo imaging to ensure that the image disparity calculation is effective.

[0013] Furthermore, when the information acquisition module acquires the distribution of surrounding complementary targets, it simultaneously collects the relative positional relationship between the complementary targets and the obstructions to form a spatial obstacle distribution map.

[0014] Furthermore, when the flight control module controls the drone for inspection: The system compares the deviation between the actual trajectory and the optimal path in real time. When the deviation exceeds the allowable range, the flight speed and turning angle are adjusted synchronously to correct the trajectory. Combined with the image clarity parameters fed back by the binocular vision acquisition module, the inspection altitude is dynamically adjusted. When the clarity is lower than the recognizable threshold, the flight altitude is reduced to improve the image resolution and ensure accurate extraction of target features.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention solves the problem of adaptive planning for different types of complementary target sub-regions. It can generate targeted paths based on the distribution density gradient and direction of the first type of target sub-region, and the altitude difference and horizontal distance between the second type of target sub-region and the UAV. Secondly, it solves the problem of balancing comprehensive coverage of complementary targets with ensuring flight stability. Through classification planning, it helps to ensure complete coverage of both types of target sub-regions while minimizing the curvature of the flight trajectory of the second type of target sub-region, thus ensuring smooth flight. Thirdly, it solves the problem of the continuity of multi-sub-region path integration. By integrating the paths of the two types of sub-regions, it eliminates connection conflicts and makes the path transition between different sub-regions natural, which helps to avoid repeated flights or missed observation points. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system framework structure of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This application provides a binocular vision-based unmanned aerial vehicle (UAV) inspection and control system, including: The binocular vision acquisition module is used to acquire images of the target to be inspected when the UAV is inspecting along a preset route, detect the degree of occlusion, and extract the target features and distribution patterns. The information acquisition module is used to acquire the target location, the distribution of surrounding complementary targets, the UAV pose and kinematic parameters when the occlusion degree of the detected target point exceeds the set value. The strategy selection module, based on the acquired information, identifies the occlusion type of the target point and the visible angle range of the complementary target point, divides the detection sub-region containing the first / second type of complementary targets, and selects the corresponding inspection strategy from the preset strategy library; The path planning module is used to call the model according to the strategy, generate a path for the first type of target sub-region based on the distribution density gradient and direction; generate an initial path for the second type of target sub-region based on the height difference with the UAV, combine horizontal distance correction to minimize the trajectory curvature, and integrate to obtain the optimal path. The flight control module is used to control the UAV to inspect complementary targets along the optimal path and complete the full observation of the originally obscured targets.

[0019] The binocular vision acquisition module refers to a module that uses two cameras in different positions to simulate human binocular vision and realize the perception of the three-dimensional structure and depth information of the target. Specifically, it can be implemented using various algorithms such as stereo matching algorithm and disparity calculation algorithm, and is used to accurately acquire the image of the target to be inspected and detect the degree of occlusion.

[0020] Among them, the target to be inspected refers to the specific object that needs to be inspected and monitored in the inspection task. Specifically, it can be extracted and located from the inspection scene through target recognition algorithms (such as target detection algorithms based on deep learning) to identify the core objects of the drone inspection, such as power poles and communication base stations.

[0021] Among them, occlusion and degree refer to the phenomenon that mutual occlusion between objects causes part of the target area to be invisible and the indicators for quantifying this phenomenon. Specifically, image segmentation algorithms, region overlap calculation algorithms, etc. can be used to determine and quantify it, which is used to evaluate the occlusion state of the target and provide a basis for subsequent strategy adjustments.

[0022] The information acquisition module refers to the module that automatically collects relevant key information when the target occlusion exceeds the limit. Specifically, it can use multi-sensor data fusion algorithms (such as Kalman filtering algorithm) to integrate information such as target position and UAV pose, so as to provide comprehensive data support for strategy selection.

[0023] Among them, complementary targets in the surrounding area refer to targets that are spatially related to the occluded target and can assist in observing the occluded part from other perspectives. Specifically, spatial correlation analysis algorithms and view coverage calculation algorithms can be used to identify them, so as to provide supplementary perspectives for the complete observation of the occluded target.

[0024] Among them, UAV pose refers to the UAV's position coordinates and attitude angles (pitch angle, roll angle, and heading angle) in space. Specifically, it can be calculated in real time using inertial navigation algorithms, GPS and IMU fusion algorithms, etc., to accurately describe the UAV's spatial state.

[0025] The strategy selection module refers to the module that matches and selects appropriate inspection strategies based on the acquired information. Specifically, it can be implemented using rule-based reasoning algorithms, case-based reasoning algorithms, etc., to quickly determine response solutions for different occlusion types and target distributions.

[0026] Among them, the first / second type of complementary targets refers to two types of auxiliary observation targets divided according to distribution characteristics and observation attributes. Specifically, clustering algorithms (such as DBSCAN) and feature classification algorithms (such as decision tree algorithms) can be used to divide them to achieve differentiated path planning. The first type corresponds to targets with density gradient and direction characteristics, and the second type corresponds to targets with height difference.

[0027] The preset strategy library refers to a set of inspection strategies preset for different occlusion scenarios and complementary target types. Specifically, it can be implemented by combining a rule base with a strategy model trained by reinforcement learning, including close-range surround strategy (suitable for small-scale occlusion), height adjustment strategy (suitable for vertical occlusion), feature association priority strategy (suitable for multi-target association scenarios), etc., to quickly match the optimal inspection scheme.

[0028] Among them, the path planning model refers to the mathematical model used to generate the optimal inspection path, which can specifically employ Bézier curve fitting (to ensure path smoothness) or A... The algorithm (which guarantees path optimality) combines target observation angle constraints and UAV kinematic limitations (such as maximum turning angular velocity and minimum flight altitude) to calculate collision-free and efficient inspection trajectories.

[0029] The optimal path refers to the UAV flight trajectory that comprehensively considers coverage integrity, flight efficiency, and stability. Specifically, it can be generated using multi-objective optimization algorithms (such as genetic algorithms) and trajectory smoothing algorithms to ensure that the UAV completes the inspection task in the optimal way.

[0030] The flight control module refers to the module that controls the movement state of the UAV according to the planned path. Specifically, it can be implemented using PID control algorithms, model predictive control algorithms, etc., to accurately control the UAV to fly along the optimal path and complete the inspection operation.

[0031] The innovations of this application are as follows: First, it solves the problem of adaptive planning for different types of complementary target sub-regions, generating targeted paths based on the distribution density gradient and orientation of the first type of target sub-region, and the altitude difference and horizontal distance between the second type of target sub-region and the UAV. Second, it solves the problem of balancing comprehensive coverage of complementary targets with ensuring flight stability. Through classification planning, it helps to ensure complete coverage of both types of target sub-regions while minimizing the curvature of the flight trajectory of the second type of target sub-region, ensuring smooth flight. Third, it solves the problem of the continuity of multi-sub-region path integration. By integrating the paths of the two types of sub-regions, it eliminates connection conflicts and makes the path transition between different sub-regions natural, helping to avoid duplicate flights or missed observation points. Furthermore, this technical solution also solves the problem that targets to be inspected cannot be fully observed due to obstruction during UAV inspection. By detecting obstruction, acquiring information, selecting strategies, planning paths, and controlling flight, it achieves comprehensive observation of the originally obstructed targets, making up for the shortcomings of traditional inspection systems, improving the comprehensiveness and accuracy of UAV inspection, and providing reliable data for subsequent detection and evaluation; it also optimizes the inspection path and improves inspection efficiency; the system can adapt to different obstruction conditions and target distributions, has strong flexibility and adaptability, and expands the application scope of UAV inspection.

[0032] The working principle of this application is as follows: First, a binocular camera is installed on the drone. While flying along a preset route, the two cameras simultaneously capture images of the target to be inspected. The images are processed using an image matching algorithm to calculate the target depth information to determine the degree of occlusion; target features are extracted using a feature extraction algorithm, and distribution patterns are summarized by analyzing the location information of multiple targets. Secondly, when the occlusion level exceeds the threshold, the drone's position is obtained via GPS and its pose is determined by the attitude sensor; the position of the target and its surrounding complementary targets is measured using a range sensor to determine their distribution; kinematic parameters are obtained through speed sensors and other means, and the information is then compiled and organized. Next, after receiving the information, the image recognition algorithm is used to identify the occlusion type; the visible angle range is calculated based on the complementary target position and the UAV pose; the detection sub-region is divided according to the complementary target characteristics, and a suitable inspection strategy is selected from the preset strategy library. Next, the path planning model is invoked according to the selected strategy. For the first type of target sub-region, a path is generated using gradient descent or path search algorithm based on the distribution density gradient and direction. For the second type of target sub-region, an initial path is generated by sorting the paths according to their height difference with the UAV. Then, the path is corrected by curve fitting and optimization algorithms based on the horizontal distance to minimize the trajectory curvature. Finally, the two types of paths are integrated to obtain the optimal path. Finally, the optimal path is received and converted into control commands. The flight control system controls the UAV to fly along the optimal path, inspect complementary targets, and receive feedback from the binocular vision acquisition module in real time to ensure that complete information of the original occluded targets is obtained.

[0033] In some embodiments of this application, traditional UAV inspection path planning generally adopts a uniform strategy for all target points without distinguishing the spatial distribution characteristics and type differences of the targets. This may result in two problems: first, repeated round trips may occur when planning paths for densely distributed target points, increasing the invalid flight distance; second, when planning paths for target points with large altitude differences, the flight altitude may be frequently adjusted, resulting in excessive trajectory curvature. This not only reduces inspection efficiency but also shortens the service life of the equipment due to frequent changes in direction and focus.

[0034] To address this, this application further proposes a path planning model that, based on the selected inspection strategy, considers the UAV flight trajectory, binocular vision focal length adjustment, and spatial correlation of target points. For the sub-region containing the first type of complementary target points, the model analyzes the spatial distribution density gradient of each target and generates a movement path within the sub-region in order from high-density areas to low-density areas, with the path direction consistent with the main distribution direction of the targets. For the sub-region containing the second type of complementary target points, the model calculates the relative height difference between each target and the current position of the UAV, generates an initial path in order from smallest to largest height difference, and then smooths the path by combining the horizontal distance between targets to minimize the curvature of the flight trajectory between adjacent targets. By integrating the movement paths of each sub-region, the optimal inspection sequence and flight path of complementary target points are obtained.

[0035] The specific implementation process of this application is as follows: First, for the sub-region where the first type of complementary target points (such as tower accessories in power line inspection: surge arresters, terminals) are generated, a density gradient is generated through spatial distribution clustering analysis. Target 3D coordinate data is collected (based on GPS and LiDAR joint positioning, accuracy ±0.5m). The DBSCAN clustering algorithm is used, with the Euclidean distance between targets as the basis (setting a neighborhood radius of 2m), to identify the degree of clustering: targets less than 2m apart are classified as dense clusters, and those greater than 5m are classified as sparse points. The distribution interval of targets within a cluster is calculated (e.g., the average interval between targets in a dense cluster is 0.8m, and the average interval in a sparse area is 6m). A density gradient is constructed based on the interval difference (the smaller the interval, the higher the gradient value), reflecting the density of the distribution without calculating the quantity per unit area. Secondly, paths are generated from high to low gradient (dense to sparse), and the path direction is consistent with the main distribution direction of the targets. If the targets are linearly distributed along the line direction (such as the accessories on both sides of the tower arranged along the conductor), the main distribution direction (such as azimuth angle 30°) is extracted through principal component analysis, and the path extends along the 30° direction. When passing through dense clusters, a serpentine trajectory is adopted (turning angle ≤ 30°), which helps to ensure that each target is within the binocular visual focal length range (focal length locked at 50mm, imaging distance 3-8m). After transitioning to the sparse region, the flight switches to a straight line to reduce turning energy consumption.

[0036] Furthermore, for the path generation of the second type of complementary target points (such as insulator strings and anti-vibration hammers in power line inspection), during the initial path generation, the relative height difference is first calculated by using the three-dimensional coordinate difference between the current position of the UAV (via GNSS positioning) and the target point (e.g., target A is at an altitude of 100m, the current UAV is at 95m, the height difference is 5m), and then sorted by height difference from smallest to largest (inspecting targets at 5m first, then 10m, and finally 15m).

[0037] During the trajectory smoothing correction stage, the horizontal distance between adjacent targets is calculated (e.g., the horizontal projection distance between targets A and B is 8m). The trajectory is fitted using a Bezier curve to ensure that the turning radius is greater than or equal to the minimum turning radius of the UAV (set to 5m) and to ensure that the trajectory curvature is minimized (curvature value ≤ 0.2 / m).

[0038] If the corrected path exceeds the equipment's performance (e.g., the climb rate exceeds 3m / s), the path will be split into multiple segments (e.g., the 15m height difference will be split into 5m + 10m) to complete the inspection step by step.

[0039] For the connecting sections of the two types of sub-region paths (such as from the tower attachment to the insulator string), calculate the coordinates of the transition point (select the common point closest to the two regions) to ensure a smooth transition of flight speed at the connection point (from 5m / s to 3m / s and then up).

[0040] The particle swarm optimization algorithm is used to select the path with the shortest total flight distance and satisfactory image quality from the set of feasible paths (with a sharpness threshold of 0.8, calculated using an image sharpness evaluation function).

[0041] Output the optimal inspection sequence (e.g., dense area accessories -- sparse area accessories -- low-height insulators -- high-height insulators) and the corresponding flight parameters (speed, altitude, focal length).

[0042] The innovation of this application lies in solving the problem of insufficient adaptation to target distribution characteristics in traditional path planning. For densely distributed targets, cluster analysis and orientation adaptation are used to avoid missed detections and duplicate inspections, thereby improving inspection coverage. For targets with varying altitudes, altitude difference sorting and trajectory smoothing are used to reduce flight energy consumption and equipment wear, thereby extending flight time. The entire process combines binocular vision focal length and equipment performance constraints to ensure imaging quality and provide clear data for subsequent defect identification.

[0043] In some embodiments of this application, traditional UAV visual acquisition modules often fail to address target occlusion issues during inspections, leading to subsequent misjudgments (such as mistakenly classifying occluded bolts as missing). Furthermore, the extraction of target distribution patterns relies solely on two-dimensional images, ignoring three-dimensional spatial features, making it difficult for path planning modules to accurately determine distribution directions.

[0044] In response, this application further proposes to capture the three-dimensional image information of the target under inspection in real time using a binocular camera, use an edge detection algorithm to identify the occlusion contour, calculate the proportion of the occlusion area to the total area of ​​the target to determine the degree of occlusion, and extract the three-dimensional spatial coordinates, geometric dimensions and arrangement direction features of the target as distribution pattern information, wherein the arrangement direction features are obtained by detecting the main distribution axis of the target through Hough transform.

[0045] Specifically, the binocular camera acquires the left and right views of the target under inspection through synchronous exposure, calculates the pixel disparity using a stereo matching algorithm (such as the SGBM algorithm), and generates a three-dimensional point cloud model of the target—that is, the three-dimensional spatial coordinates (X, Y, Z) corresponding to each pixel. The Canny edge detection algorithm is used to identify the occlusion contours of the target. The 3D point cloud model is converted to grayscale, and the pixel gradient changes of the target edges are extracted (gradient values ​​> 100 are considered edges). The edge differences between the complete target model (such as the 3D model of a standard insulator string) and the actual acquired model are compared, and the difference area is the occlusion contour. The proportion of the number of pixels covered by the occlusion contour to the total number of pixels of the target is calculated to determine the degree of occlusion: < 10% is slight occlusion, 10%-30% is moderate occlusion, and > 30% is severe occlusion. This data is synchronized to the path planning module (severe occlusion may trigger secondary inspection).

[0046] When extracting the three-dimensional spatial coordinates of the target, a local coordinate system is established with the current position of the UAV as the origin. The three-dimensional coordinates of each target are directly obtained through the point cloud model (e.g., the coordinates of the pole attachment are (5.2m, 3.1m, 15.8m)), providing basic data for distance calculation in path planning. Geometric dimension extraction is based on distance calculation from a 3D point cloud: the target point cloud is clustered and segmented (e.g., separating the point cloud of an insulator string from the background); the spatial distances of key target components are calculated, such as the length of the insulator string (distance between its two endpoints) and the diameter of the bolt (circumscribed circle diameter), with an accuracy of ±1mm, providing dimensional basis for defect detection (e.g., insulator breakage). The arrangement direction characteristics are obtained by detecting the target's main distribution axis through Hough transform: the target point cloud is projected to obtain a point set on a two-dimensional plane (e.g., projection of attachments along the tower direction). The Hough transform algorithm is applied to convert the pixel coordinates in the point set into polar coordinates. The cumulative weights corresponding to angle θ (angle range 0-180°) are calculated, and the angle θ corresponding to the peak weight is the direction of the main distribution axis of the target. For example, the Hough transform results of pole and tower accessories in power line inspection show that the weight is highest when θ=30°, indicating that its main distribution axis is 30°. This data is directly used to set the parameters in the path planning module to make the path direction consistent with the target distribution direction. Data such as occlusion level, three-dimensional spatial coordinates, geometric dimensions, and main distribution axis are packaged and transmitted in real time to the path planning module via a 4G / 5G module, providing accurate data support for path optimization.

[0047] The innovation of this application lies in the fact that by combining 3D point cloud with edge detection, it can accurately identify occlusion contours and quantify the degree of occlusion, avoiding misjudgment of defects caused by occlusion, and providing reliable and effective image areas for subsequent analysis. Based on the distribution pattern information extracted by 3D coordinates and Hough transform (especially the main distribution axis), it provides an accurate basis for the path planning module to generate a path consistent with the target direction, thereby improving the adaptability of the inspection trajectory and the target distribution.

[0048] In some embodiments of this application, when planning a path for a first type of target sub-region (such as pole attachments in power line inspection), if there is a local conflict between the distribution density gradient and the main direction, the basic path generated according to the main direction may deviate from the high-density area, resulting in missed detection; forced coverage may also affect the stability of the UAV due to excessive turning, resulting in a sudden change in the subsequent target perspective and affecting the continuity of the image.

[0049] In response, this application further proposes that when initially generating the basic path along the main direction, conflict segments in the path that deviate from the local high-density cluster by a set value should be marked. By using information on the distribution patterns of the target (three-dimensional coordinates and arrangement direction), the spatial distance between the path and local high-density clusters (such as densely bolted areas on the side of a tower) can be calculated using the Euclidean distance formula. When this distance exceeds a set value (such as 1.5m), this section of the path is marked as a conflict segment—meaning that the basic path cannot effectively cover the high-density cluster.

[0050] For conflict segments, branch paths parallel to the main direction are generated to cover high-density clusters. The starting and ending points of the branch paths fall on the edges of the high-density clusters to ensure coverage of all targets within the clusters. The branch paths are connected to the base paths through arc-shaped transition segments, and the rate of change of the turning angle of the transition segments is calculated. It should be noted that the radius of curvature of the transition section is set according to the minimum turning radius of the UAV (e.g., 5m), and the rate of change of the turning angle of the transition section (the change in the turning angle per unit time) is calculated. For example, if the base path is vertical and the branch path is parallel to the vertical, covering the high-density cluster on the right, the transition section turns from the base path to the branch path in an arc to the right, and the initial calculation of the rate of change of the turning angle is 25° / s².

[0051] If the rate of change of the steering angle exceeds the attitude stability threshold of the UAV, it indicates that a single overly abrupt turn may cause the fuselage to shake. In this case, the branch path length should be shortened and the number of transition segments increased. For example, instead of covering the entire high-density cluster, the transition segment should be changed to cover the core area. One transition segment should be split into two consecutive arc segments, and the rate of change of the steering angle should be recalculated until it meets the threshold requirement. For example, the original transition segment had a steering angle of 30° and a time of 1s, with a rate of change of 30° / s². After splitting, each transition segment has a steering angle of 15° and a time of 1s, and the rate of change is reduced to 15° / s², which meets the threshold requirement.

[0052] After each round of adjustment, verify the deviation of the overall path from the target main direction (e.g., an angle ≤ 5° is considered low deviation). If the deviation does not decrease after two consecutive rounds of adjustment (e.g., it remains at 8°), it indicates that the interference of the branch path on the main direction cannot be resolved by shortening its length. In this case, divide the local high-density cluster into an independent sub-region that still belongs to the first type of target sub-region and plan a separate path. Repeat the verification until the conflict is resolved. For example, set the dense bolt area on the side of the tower as a sub-region, plan the path along its own distribution direction (diagonal), and connect it with the original foundation path through multiple small-angle transition sections.

[0053] Furthermore, for the paths in the independent subdivided areas, the deviation from the main direction and the rate of change of the turning angle are checked again until the conflict is resolved (e.g., the path covers all high-density targets, the rate of change of the turning angle is ≤20° / s², and the deviation is ≤5°).

[0054] The innovation of this application lies in the fact that, through the design of branch paths and arc transition sections, it can cover local high-density targets while ensuring the main direction, avoiding missed or repeated inspections. The dynamic adjustment of the turning angle change rate helps to ensure that the UAV maintains a stable attitude in complex paths, reducing imaging blur caused by jitter. The independent subdivision mechanism provides a flexible solution for specially distributed high-density clusters, which helps to avoid the overall path deviating too much from the main direction. This process not only ensures the inspection coverage rate, but also takes into account the flight safety of the UAV, improving inspection efficiency and data reliability. Furthermore, it achieves synergistic optimization of macroscopic path stability and microscopic target coverage, rather than simple parameter balancing. It transforms the abstract requirement of image coherence into operable iterative verification standards such as continuous judgment of viewpoint change features and local path adaptation within segments, establishing an implicit correlation between path geometric parameters and visual acquisition quality. This cross-dimensional parameter mapping relationship breaks the traditional limitation of path planning focusing only on spatial coverage. It not only ensures that the overall path does not deviate from the main target direction (reducing macroscopic viewpoint abrupt changes) but also fully covers local high-density clusters (avoiding the omission of microscopic targets). Simultaneously, the design of transition sections ensures continuous and stable viewpoint changes.

[0055] In some embodiments of this application, for the second type of target sub-region, the sorting of altitude differences and the correction of horizontal distance may cause implicit path redundancy. For example, two targets with small altitude differences but extremely large horizontal distances will be preferentially connected in series according to the altitude difference sorting, causing the flight trajectory to detour significantly in the horizontal direction, which will increase energy consumption and inspection time. How to solve this discrepancy between local optima and global efficiency is a difficult problem.

[0056] In response, this application further proposes that after the initial path is generated, the ratio of the horizontal distance to the height difference of each pair of adjacent targets is calculated, and abnormal target pairs whose ratio exceeds a set threshold are marked. For example, if the horizontal distance between target A and target B is 25m and the height difference is 2m, the ratio is 12.5. If the threshold is set to 10, they will be marked as an abnormal target pair.

[0057] For marked abnormal target pairs, other targets within a certain range around them are retrieved as potential transition points. A branch path scheme connecting the two targets through the transition point is constructed, and the energy consumption difference between this scheme and the original straight path is calculated. That is, flying from target A to the transition point and then from the transition point to target B, the energy consumption of this scheme (e.g., total flight distance 30m, power consumption 1.5Wh) is calculated using a three-dimensional path planning algorithm, and the energy consumption difference between this scheme and the original straight path (flight distance 25m, power consumption 1.8Wh) is calculated.

[0058] If the energy consumption of the branch path is lower than that of the original path, the branch path is used to replace the original connection method, the transition point is included in the path sequence, and the ratio of all adjacent targets in the new sequence is recalculated. The marking and replacement steps are repeated to ensure that there are no abnormal target pairs in the new sequence.

[0059] After each round of replacement, verify whether the total inspection time of the overall path has been shortened. If the total time does not improve after two consecutive rounds of replacement, stop the loop and retain the current optimized path. After each round of replacement, calculate the total inspection time of the overall path based on the drone's flight speed (e.g., horizontal speed 8 m / s, climb speed 3 m / s) and the total path length. If the total time does not improve after two consecutive rounds of replacement (e.g., it remains at 5 minutes), stop the loop and retain the current optimized path to avoid invalid iterations.

[0060] Finally, the optimized path undergoes a turnaround rate check. If the number of turnarounds in the same area exceeds a set value, a spatial partitioning algorithm is invoked to cluster the targets by horizontal distance. The number of turnarounds by the drone in the same area (e.g., a radius of 50m) is counted. If this exceeds a set value (e.g., 3 times), a spatial partitioning algorithm (e.g., K-means clustering) is invoked to cluster the targets by horizontal distance. For example, targets with a horizontal distance of less than 15m are grouped into the same cluster, and priority is given to sorting by altitude difference within each cluster, achieving a balance between local compactness and global efficiency.

[0061] The innovation of this application lies in marking abnormal target pairs and constructing branch paths, which reduces invalid lateral flight and lowers energy consumption while ensuring the sorting logic based on altitude difference. Through multi-round cyclical collaborative judgment, it can both retain the initial efficiency of altitude difference sorting and eliminate the implicit redundancy caused by direct connection of distant low-difference targets. At the same time, by combining turnaround rate detection and spatial partitioning clustering, it can balance the compactness of local paths with global inspection efficiency, reduce repeated turnarounds in the same area, and help solve the core problem of the divergence between local optima and global efficiency. This process not only improves the energy economy of the path but also shortens the total inspection time, making the path planning of the second type of target sub-region more efficient and reasonable.

[0062] In some embodiments of this application, when integrating the paths of a first type of target sub-region (such as tower attachments) and a second type of target sub-region (such as insulator strings), the difference in the orientation of the two types of paths may cause a sharp change in direction at the junction, making it difficult for the UAV to adjust its attitude, and even exacerbating the trajectory deviation due to airflow disturbances. For example, in power line inspection, when the attachment path along the vertical direction of the tower connects with the insulator path along the horizontal direction of the conductor, if the angle of the change in direction reaches 90°, the UAV may cause blurred images due to excessive turning, and this effect will be more significant in strong airflow environments.

[0063] In response, this application further proposes that after initial integration, the trajectory tangent direction at the junction of the two types of paths is obtained (e.g., the tangent direction of the attached path is vertically upward, and the tangent direction of the insulator path is horizontally to the right and the current motion vector of the UAV (speed magnitude and direction), the direction change angle is calculated by the vector angle formula (90° in the above example), and the transition segment where the direction change angle exceeds the set value (e.g., 45° in power inspection); For the marked transition segment, a virtual guide point is introduced (the angle bisector direction is 45° in the example above). This guide point is located on the angle bisector of the angle between the tangents of the two types of paths, at a certain distance from the connection point, and a transition path connected by the guide point is constructed. The initial distance between the virtual guide point and the connection point is set to 1.5 times the minimum turning radius of the UAV (e.g., 7.5m when the turning radius is 5m). Using the virtual guide point as the intermediate node, a transition path is constructed from the end point of the first type of path through the virtual guide point to the starting point of the second type of path, thus transforming the original polygonal trajectory into a smooth curve.

[0064] Calculate the rate of curvature change of the transition path. If the rate of curvature change exceeds the threshold of the UAV's attitude adjustment capability, adjust the position of the guide point, increase its distance from the connection point, reconstruct the transition path, and repeat the calculation until the rate of curvature change meets the standard. The method for calculating the rate of change of curvature of the transition path is as follows: fit the transition path with a cubic spline curve to obtain a coordinate equation with arc length as a parameter, take the second derivative of the coordinate equation, calculate the curvature of each point on the path (reflecting the degree of curvature of the curve), and calculate the ratio of the curvature difference between two adjacent points to the arc length difference, which is the rate of change of curvature of that segment (reflecting how fast the degree of curvature changes). For example, if the arc length difference between two points on the transition path is 2 meters, the curvatures are 0.2 / m and 0.3 / m respectively, and the rate of change of curvature is (0.3-0.2) / 2=0.05 / m². If the rate of curvature change exceeds the drone's attitude adjustment capability threshold (e.g., 0.05 / m²), the distance between the guide point and the connection point is increased (e.g., from 7.5m to 10m), and the transition path is refitted. For example, if the original path had a rate of curvature change of 0.07 / m² due to the guide point being too close, adjusting the distance reduces it to 0.04 / m², which meets the threshold requirement.

[0065] By combining real-time airflow disturbance data (such as a crosswind of 6 m / s detected by a wind speed sensor, corresponding to a set disturbance intensity level of 3), the transition path is optimized. When the airflow disturbance intensity exceeds the set level (such as level 2), the number of guiding points is increased (such as from 1 to 3), so that the transition path is a multi-segment smooth curve, and the curvature center of each segment of the curve points in the opposite direction of the airflow disturbance; (when the crosswind comes from the left, the curvature center shifts to the right), thus offsetting the impact of the airflow on the trajectory.

[0066] After each round of optimization, the connection accuracy between the transition section and the two types of paths is checked, that is, the deviation value between the endpoint of the transition path and the original connection point. If the deviation value does not decrease after two consecutive optimizations, a dynamic segmentation strategy that is only for cross-type path connection scenarios is activated. The transition section is divided into different sections according to the intensity of airflow disturbance. Each section adopts an appropriate curve type. The process is repeated until the direction change angle, curvature change rate and connection accuracy all meet the requirements.

[0067] Furthermore, after each round of optimization, the deviation between the endpoint of the transition path and the original connection point—that is, the straight-line distance between the two points in three-dimensional space—is detected. The deviation between the endpoint of the transition path and the original connection point is calculated using GPS positioning data (e.g., the allowable deviation is ≤0.5m). If the deviation does not decrease after two consecutive optimizations (e.g., it remains at 0.8m), a dynamic segmentation strategy for cross-type path connections is activated: the transition segment is divided into different sections according to the intensity of airflow disturbance (e.g., strong disturbance section, medium disturbance section, weak disturbance section). The strong disturbance section uses a circular arc curve with smaller curvature, and the weak disturbance section uses a gentle spline curve.

[0068] The adjusted transition section is repeatedly verified until the direction change angle (e.g., reduced to 30°), curvature change rate (e.g., 0.03 / m²), and connection accuracy (e.g., deviation 0.3m) all meet the requirements, and finally a smooth transition integrated path is output.

[0069] The innovation of this application lies in transforming abrupt directional changes into smooth transitions through virtual guide points and curvature adjustments. This helps ensure the stability of the UAV's attitude, avoids blurry images, and enhances the trajectory's anti-interference capability under complex weather conditions by dynamically optimizing the path in conjunction with airflow disturbances. The dynamic segmentation strategy balances local connection accuracy with global path smoothness, helping to ensure the continuity of cross-type area inspections. It solves the problems of abrupt directional changes, poor airflow adaptability, and low connection accuracy in the transition section when integrating paths of two types of sub-regions, and maintains the stability of the UAV in the transition section under environmental interference. This process makes the UAV switch between two types of target areas more smoothly, further optimizing the overall inspection experience.

[0070] This application further proposes to acquire stereo image pairs of the target using a binocular camera, calculate disparity maps to obtain target depth information, and employ a pixel-level occlusion detection algorithm to statistically analyze the proportion of pixels in the occluded area to the total number of pixels in the target. The occlusion is divided into three levels: mild occlusion, moderate occlusion, and severe occlusion, based on the proportion range. Mild occlusion corresponds to occlusion areas that do not affect the recognition of the target's core features, while severe occlusion corresponds to the complete obscuration of the target's core features.

[0071] Specifically, a stereo camera is used to simultaneously capture images of the target from different angles, obtaining a pair of stereo images of the target. Because the positions of the two cameras differ, the target's image position will differ in the two images; this positional difference is called parallax. Based on the principle of triangulation, the depth information of each point on the target can be obtained by calculating the parallax map. For example, in a power line inspection scenario, for insulator strings on a tower, the distance between the insulators at different positions and the camera can be clearly distinguished using the parallax map. A pixel-level occlusion detection algorithm is used to process the acquired image. This algorithm iterates through every pixel in the image, classifying pixels within the target region into occluded and unoccluded categories. By counting the number of pixels in the occluded area and calculating its proportion to the total number of target pixels, the occlusion level is determined based on a pre-defined proportion range. Specifically, when the pixel percentage of the occluded area is between 0% and 10%, it is considered mild occlusion. Although there is partial occlusion, the core features of the target (such as the shape of the insulator string and key connection parts) are still clearly identifiable and do not affect the accurate identification of the target. When the pixel percentage of the occluded area is between 10% and 50%, it is classified as moderate occlusion. The core features of the target are partially affected, but a certain degree of judgment can still be made through the remaining visible parts. When the pixel percentage of the occluded area exceeds 50%, it is defined as severe occlusion. At this time, the core features of the target are completely invisible. For example, the insulator string may be completely covered by large leaves or other objects, making it difficult to determine its original shape and state from the image. Through this set of occlusion detection processes based on binocular vision, the occlusion status of the target can be quickly and accurately determined, providing strong data support for subsequent task decisions (such as whether to readjust the shooting angle or prioritize the detection of unoccluded targets).

[0072] In some embodiments of this application, if the strategy selection module lacks a targeted strategy, there may be problems such as incomplete collection of first-type targets with slight occlusion or wasted inspection of second-type targets with heavy occlusion.

[0073] In response, this application further proposes a dense coverage strategy for combinations of mild occlusion and first-type complementary targets, with the path set to a multi-round cross-coverage mode to improve the integrity of feature acquisition; and an efficient detour strategy for combinations of severe occlusion and second-type complementary targets, with the path focusing on covering key complementary targets with the shortest distance; during strategy matching, priority is given to ensuring that the visual angle of complementary targets meets the baseline requirements of binocular stereo imaging, ensuring the effectiveness of image disparity calculation.

[0074] Specifically, the preset strategy library has two types of strategies: For the first type of target with slight occlusion (such as partially occluded tower bolts), a multi-round cross-coverage strategy is adopted. One round is taken along the main direction and another at a 30° angle, maintaining an 8m height and a 50mm focal length, improving the completeness of feature acquisition through multi-angle shooting. For the second type of target with severe occlusion (such as completely occluded insulator strings), an efficient detour strategy is adopted. Priority is given to covering key complementary targets (such as adjacent unoccluded insulators), and the path is planned according to the shortest distance, which helps avoid ineffective flights. During strategy matching, priority is given to ensuring that the visual angle of complementary targets meets the requirements of binocular visual baseline, i.e., the horizontal angle is between 30° and 60°, ensuring effective parallax calculation (parallax 50-200 pixels), providing support for subsequent depth information extraction.

[0075] This application further proposes that when the information acquisition module acquires the distribution of surrounding complementary targets, it simultaneously collects the relative positional relationship between the complementary targets and the obstructions to form a spatial obstacle distribution map.

[0076] Specifically, when the information acquisition module acquires the distribution of surrounding complementary targets (such as pole accessories and insulator strings in power line inspection), it simultaneously collects the three-dimensional coordinates of each complementary target through sensors (such as lidar and binocular cameras) to determine their spatial distribution. At the same time, it records the relative position of obstructions (such as trees and buildings) to each complementary target, including distance (e.g., obstruction is 3 meters away from the target) and orientation (e.g., obstruction is directly in front of the target). This information is integrated into a spatial obstacle distribution map, which clearly marks the distribution of complementary targets, the position of obstructions, and the relative relationship between them. This provides an intuitive spatial reference for the subsequent strategy selection module to formulate inspection paths, which helps to avoid unreasonable planning of inspection paths due to unclear obstacle information.

[0077] In some embodiments of this application, if the flight control module does not correct the trajectory deviation in real time and does not dynamically adjust the inspection altitude, it may result in: the actual trajectory deviating too far from the optimal path, causing the target to be missed; and insufficient image clarity, making it impossible to accurately extract target features, thus affecting the inspection quality.

[0078] In response, this application further proposes to compare the deviation between the actual trajectory and the optimal path in real time. When the deviation exceeds the allowable range, the flight speed and turning angle are adjusted synchronously to correct the trajectory. Combined with the image clarity parameters fed back by the binocular vision acquisition module, the inspection altitude is dynamically adjusted. When the clarity is lower than the recognizable threshold, the flight altitude is reduced to improve the image resolution and ensure accurate extraction of target features.

[0079] Specifically, when the flight control module controls the UAV for inspection, it compares the deviation between the actual flight trajectory and the optimal path in real time. When the deviation exceeds the allowable range (such as a horizontal or vertical deviation exceeding 0.5 meters), it will simultaneously adjust the flight speed (such as increasing or decreasing the speed) and the turning angle (such as increasing or decreasing the turning angle) to bring the trajectory back to the optimal path.

[0080] Simultaneously, the inspection altitude is dynamically adjusted based on the image clarity parameters fed back from the binocular vision acquisition module. When the clarity is below the recognizable threshold (e.g., the image sharpness value is below the set standard), the flight altitude is reduced (e.g., from 10 meters to 8 meters) to improve image resolution, ensuring accurate extraction of target features and guaranteeing the effective conduct of the inspection work.

[0081] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A UAV inspection and control system based on binocular vision, characterized in that, include: The binocular vision acquisition module is used to acquire images of the target to be inspected when the UAV is inspecting along a preset route, detect the degree of occlusion, and extract the target features and distribution patterns. The information acquisition module is used to acquire the target location, the distribution of surrounding complementary targets, the UAV pose and kinematic parameters when the occlusion degree of the detected target point exceeds the set value. The strategy selection module, based on the acquired information, identifies the occlusion type of the target point and the visible angle range of the complementary target point, divides the detection sub-region containing the first / second type of complementary targets, and selects the corresponding inspection strategy from the preset strategy library; The path planning module is used to call the model according to the strategy and generate a path for the first type of target sub-region based on the distribution density gradient and direction. For the second type of target sub-region, an initial path is generated by sorting the target sub-regions according to their altitude difference with the UAV. The path is then combined with horizontal distance correction to minimize the trajectory curvature and obtain the optimal path. The flight control module is used to control the UAV to inspect complementary targets along the optimal path and complete the full observation of the originally obscured targets.

2. The UAV inspection and control system based on binocular vision according to claim 1, characterized in that, The specific tasks of the path planning module are as follows: Based on the selected inspection strategy, this system invokes a corresponding path planning model that considers the UAV's flight trajectory, binocular vision focus adjustment, and spatial correlation of target points. For the sub-region containing the first type of complementary target points, it analyzes the spatial distribution density gradient of each target and generates a movement path within the sub-region in order from high-density areas to low-density areas, with the path direction consistent with the main distribution direction of the targets. For the sub-region containing the second type of complementary target points, it calculates the relative height difference between each target and the current position of the UAV, generates an initial path in order from smallest to largest height difference, and then smooths the path by combining the horizontal distance between targets to minimize the curvature of the flight trajectory between adjacent targets. By integrating the movement paths of each sub-region, the optimal inspection sequence and flight path of complementary target points are obtained.

3. The UAV inspection and control system based on binocular vision according to claim 1, characterized in that, The binocular vision acquisition module is specifically used for: The three-dimensional image information of the target under inspection is captured in real time by a binocular camera. An edge detection algorithm is used to identify the occlusion contour and calculate the proportion of the occluded area to the total area of ​​the target to determine the degree of occlusion. The three-dimensional spatial coordinates, geometric dimensions and arrangement direction features of the target are extracted as distribution pattern information. The arrangement direction features are obtained by detecting the main distribution axis of the target through Hough transform.

4. The UAV inspection and control system based on binocular vision according to claim 1, characterized in that, When the path planning module generates a path for the first type of target sub-region based on the distribution density gradient and orientation, it addresses local conflicts between the distribution density gradient and orientation, including: When initially generating the basic path along the main direction, mark the conflict segments in the path whose deviation distance from the local high-density cluster exceeds a set value; For conflict segments, branch paths parallel to the main direction are generated to cover high-density clusters. The branch paths are connected to the base paths through arc-shaped transition segments, and the rate of change of the turning angle of the transition segments is calculated. If the rate of change of steering angle exceeds the attitude stability threshold of the UAV, shorten the branch path length and increase the number of transition segments, recalculate the rate of change of steering angle until it meets the threshold requirement; After each round of adjustments, the deviation between the overall path and the target main direction is verified. If the deviation does not decrease after two consecutive rounds of adjustments, the local high-density cluster is divided into an independent sub-region that still belongs to the first type of target sub-region, and a separate path is planned. The verification is repeated until the conflict is resolved.

5. The UAV inspection and control system based on binocular vision according to claim 1, characterized in that, When the path planning module generates and corrects the initial path for the second type of target sub-regions by sorting them according to their height differences, it also includes: After the initial path is generated, the ratio of horizontal distance to height difference for each pair of adjacent targets is calculated, and abnormal target pairs with ratios exceeding a set threshold are marked. For marked abnormal target pairs, other targets within a certain range around them are retrieved as potential transition points, and branch path schemes connecting the two targets through the transition points are constructed. The energy consumption difference between this scheme and the original straight path is calculated. If the energy consumption of the branch path is lower than that of the original path, the branch path is used to replace the original connection method, the transition point is included in the path sequence, and the ratio of all adjacent targets in the new sequence is recalculated. The marking and replacement steps are repeated. After each round of replacement, it is verified whether the total inspection time of the overall path has been shortened. If the total time does not improve after two consecutive rounds of replacement, the loop is stopped, and the current optimized path is retained. Finally, the turnaround rate of the optimized path is detected. If the number of turnarounds in the same area exceeds a set value, the spatial partitioning algorithm is called to cluster the targets by horizontal distance and prioritize sorting them by height difference within the cluster to achieve a synergy between local compactness and global efficiency.

6. The UAV inspection and control system based on binocular vision according to claim 1, characterized in that, When integrating the two types of sub-region paths, the route planning module also includes: After initial integration, the trajectory tangent direction and the current motion vector of the UAV at the junction of the two types of paths are obtained, the direction change angle is calculated, and the transition segment where the direction change angle exceeds the set value is marked; For the marked transition segment, a virtual guide point is introduced. This guide point is located on the angle bisector of the angle between the tangents of the two types of paths, at a certain distance from the connection point, and a transition path connected by the guide point is constructed. Calculate the rate of curvature change of the transition path. If the rate of curvature change exceeds the threshold of the UAV's attitude adjustment capability, adjust the position of the guide point, increase its distance from the connection point, reconstruct the transition path, and repeat the calculation until the rate of curvature change meets the standard. The transition path is optimized by combining real-time airflow disturbance data. When the airflow disturbance intensity exceeds the set level, the number of guiding points is increased so that the transition path is a multi-segment smooth curve, with the curvature center of each segment pointing in the opposite direction of the airflow disturbance. After each round of optimization, the connection accuracy between the transition section and the two types of paths is checked, that is, the deviation value between the endpoint of the transition path and the original connection point. If the deviation value does not decrease after two consecutive optimizations, a dynamic segmentation strategy that is only for cross-type path connection scenarios is activated. The transition section is divided into different sections according to the intensity of airflow disturbance. Each section adopts an appropriate curve type. The process is repeated until the direction change angle, curvature change rate and connection accuracy all meet the requirements.

7. The UAV inspection and control system based on binocular vision according to claim 1, characterized in that, When the binocular vision acquisition module detects occlusion and its degree, it includes: The target stereo image pairs are acquired by a binocular camera, and the disparity map is calculated to obtain the target depth information. A pixel-level occlusion detection algorithm is adopted to count the proportion of pixels in the occluded area to the total number of pixels in the target. The proportion range is divided into three levels: mild occlusion, moderate occlusion, and severe occlusion. Mild occlusion means that the occluded area does not affect the recognition of the target's core features, while severe occlusion means that the target's core features are completely invisible.

8. The UAV inspection and control system based on binocular vision according to claim 1, characterized in that, The preset strategy library of the strategy selection module includes: For the dense coverage strategy for the combination of mild occlusion and the first type of complementary target, the path is set to a multi-round cross-coverage mode to improve the integrity of feature acquisition; for the efficient detour strategy for the combination of severe occlusion and the second type of complementary target, the path is based on covering the key complementary target with the shortest distance; when matching the strategy, priority is given to ensuring that the visual angle of the complementary target meets the baseline requirements of binocular visual stereo imaging to ensure that the image disparity calculation is effective.

9. The UAV inspection and control system based on binocular vision according to claim 1, characterized in that, When the information acquisition module acquires the distribution of surrounding complementary targets, it simultaneously collects the relative positional relationship between the complementary targets and the obstructions to form a spatial obstacle distribution map.

10. The UAV inspection and control system based on binocular vision according to claim 1, characterized in that, When the flight control module controls the drone for inspection: The system compares the deviation between the actual trajectory and the optimal path in real time. When the deviation exceeds the allowable range, the flight speed and turning angle are adjusted synchronously to correct the trajectory. Combined with the image clarity parameters fed back by the binocular vision acquisition module, the inspection altitude is dynamically adjusted. When the clarity is lower than the recognizable threshold, the flight altitude is reduced to improve the image resolution and ensure accurate extraction of target features.

Images