Unmanned aerial vehicle autonomous address selection method based on terrain evaluation and approach cost analysis

Abstract

The invention discloses an unmanned aerial vehicle autonomous address selection method based on terrain evaluation and approach cost analysis, and relates to the technical field of unmanned aerial vehicle autonomous address selection. After the unmanned aerial vehicle receives a landing instruction in an unknown environment, the geometric safety of the terrain and the cost problem of the approach path can be comprehensively considered, and the landing point selection strategy of the unmanned aerial vehicle is optimized from the global perspective. Firstly, semantic segmentation is performed on an RGB image acquired by a visual sensor through deep learning to acquire a semantic image, then point cloud data acquired by a laser radar and the semantic image are fused to acquire a three-dimensional semantic point cloud, and a terrain safety area conforming to terrain geometric feature constraints is screened out on the basis of considering semantic tags. And finally, calculating the cost of the linear approach path of each candidate landing point in the terrain safety area through traversal, and screening out the landing point with the optimal comprehensive cost, thereby improving the accuracy and practicability of terrain safety area identification. According to the method, an unreachable or high-risk area is avoided, and the global point selection optimization effect is improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous location selection technology for unmanned aerial vehicles (UAVs), specifically a method for autonomous location selection of UAVs based on terrain assessment and approach cost analysis. Background Technology

[0002] In the context of information-based and intelligent warfare, unmanned equipment has become an important component of new combat platforms. With its low cost, multi-functionality, and flexibility, unmanned aerial vehicles (UAVs) are gradually being integrated into the entire weapon system, playing an irreplaceable role in urban warfare, border defense, reconnaissance, and surveillance. UAV landing, as a crucial component of the takeoff and landing system, is fundamental to performing various civilian and military missions. Safe landing effectively avoids accidents in normal or emergency situations. The prerequisite for safe UAV landing is selecting a reliable landing area or identifying an area capable of delivering the payload, which is critical for the UAV system.

[0003] Currently, research on autonomous landing for unmanned aerial vehicles (UAVs) largely focuses on using sensors such as vision and lidar to acquire environmental information, identifying potential landing areas through image recognition or terrain reconstruction techniques, and typically using indicators such as terrain flatness, safety category, and open space as criteria for selection. For example, a Chinese invention patent with publication number "CN113359810B" entitled "A Method for Identifying UAV Landing Areas Based on Multiple Sensors" provides a method for identifying UAV landing areas based on multiple sensors. This patent mainly includes the following steps: First, feature extraction and semantic segmentation are performed on images captured by an airborne camera to identify possible landing areas; then, after the UAV flies to low altitude, a three-dimensional point cloud map is constructed by combining sensors such as lidar, IMU, and GPS, and further semantic segmentation and environmental modeling are performed on the point cloud data through deep learning to generate a three-dimensional semantic map; finally, geometric methods are used to evaluate the terrain to extract the location information of areas that can be used for safe landing. However, this method only focuses on the identification and location of landable areas on the ground, without cost modeling of the approach path of the UAV from its current location to the target landing area. This may result in the selected area being landable, but the flight path being high-risk or inaccessible.

[0004] However, existing methods neglect the approach cost of the UAV's flight path from its current location to the candidate landing point. In practical applications, even if a region has favorable terrain for landing, the presence of numerous obstacles or an excessively long, high-risk, or energy-intensive flight path may render it a less than optimal landing choice. To address this issue, UAVs need to possess both terrain assessment and approach cost analysis capabilities. This would ensure the safety of the selected landing area while minimizing approach risks, thereby enhancing the UAV's autonomous landing capability and environmental adaptability in actual flight missions. Summary of the Invention

[0005] To address the above problems, this invention proposes an autonomous landing method for unmanned aerial vehicles (UAVs) based on terrain assessment and approach cost analysis. This method enables UAVs to comprehensively consider the geometric safety of the terrain and the cost of the approach path after receiving a landing command in an unknown environment, thereby optimizing the UAV's landing point selection strategy from a global perspective and improving its decision-making ability to autonomously select landing points in unknown environments.

[0006] The technical solution of this invention is as follows: the drone is equipped with an onboard visual sensor, IMU, lidar, onboard computer, and flight controller. After receiving a landing command, the drone autonomously selects a landing site according to the following steps:

[0007] S1: RGB images are acquired in real time through a visual sensor and transmitted to an onboard computer. The acquired RGB images are semantically segmented through deep learning, and different colors are assigned to different categories to distinguish them, thus obtaining semantic images.

[0008] S2: Use the information obtained by the IMU to preprocess the point cloud data collected by the lidar, and combine it with the semantic image in step S1 to achieve information fusion and obtain three-dimensional point cloud data with semantic labels, i.e. semantic point cloud data.

[0009] S3: Perform geometric feature extraction and semantic classification fusion analysis on semantic point cloud data, calculate terrain safety distance, and identify suitable terrain safety areas for landing based on semantic labels;

[0010] S4: Iterate through and calculate the cost of the linear approach path for each candidate landing point within the terrain safety area. Taking into account both the terrain safety distance and the cost of the approach path, select the best landing point through a weighted average.

[0011] S5: Sends the location information of the selected optimal landing point to the flight controller. The flight controller then precisely guides the drone to autonomously land at the target landing point by adjusting the drone's flight attitude in real time.

[0012] The specific method for step S1 is as follows:

[0013] Step S1.1: Training the semantic segmentation model;

[0014] Aerial images were selected as the training dataset. Data augmentation strategies were used to improve the generalization ability of the semantic segmentation model. Image pixels were divided into six semantic categories: roads, buildings, cars, trees, grass, and water, and assigned color labels. Finally, a semantic segmentation model with high segmentation accuracy and strong semantic recognition ability was trained.

[0015] Step S1.2: Deployment of the semantic segmentation model and image processing;

[0016] The trained semantic segmentation model is deployed to an onboard computer. RGB images acquired by the visual sensor are transmitted to the onboard computer via an interface. The semantic segmentation model is then used to segment the RGB images and output images with semantic labels.

[0017] The specific method for step S2 is as follows:

[0018] Step S2.1: Use software time synchronization to align the vision sensor, IMU and LiDAR using timestamps or ROS time.

[0019] The external parameters of the vision sensor and LiDAR are calibrated using calibration tools to obtain the external parameter matrix. The point cloud information of the LiDAR is first projected into the coordinate system of the vision sensor, and then projected into the image coordinate system to achieve spatial alignment of multi-source data.

[0020] Among them, [x c, y c ,z c ] T Let R be the 3D position of the point in the visual sensor coordinate system, R be the rotation matrix of the LiDAR coordinate system relative to the visual sensor coordinate system, and t be the translation vector of the LiDAR coordinate system relative to the visual sensor coordinate system. The values ​​of R and t depend on the relative mounting positions of the LiDAR and the visual sensor. [x1, y1, z1] T K represents the 3D position in the point cloud lidar coordinate system, and K is the intrinsic parameter matrix of the camera.

[0021] Step S2.2: Based on the high-frequency attitude information provided by the IMU, perform distortion correction processing on the lidar;

[0022] Step S2.3: Filter the point cloud data to reduce computational complexity and save computing resources of the airborne computer;

[0023] Step S2.4: Based on the extrinsic matrix of the visual sensor and the lidar, the semantic information in the semantic image is fused into the point cloud information to obtain the semantic point cloud;

[0024] The point cloud is projected onto the image plane using an extrinsic parameter matrix. The semantic information of the corresponding pixels is obtained by nearest neighbor search, and this information is mapped back to the point cloud to generate a semantic point cloud. This achieves the fusion of the image and the point cloud and obtains semantic point cloud data.

[0025] The specific method for step S3 is as follows:

[0026] Step S3.1: The semantic point cloud is meshed according to a predetermined resolution, and the three-dimensional semantic point cloud is converted into a two-dimensional elevation grid map and semantic grid map, wherein each cell in the grid contains a height value or semantic information.

[0027] Step S3.2: Calculate the terrain slope based on the elevation grid map. Use principal component analysis to calculate the surface normal of each grid and its neighboring grids in the z-axis direction. Then, obtain the slope of the grid relative to the horizontal plane using the following formula:

[0028] Where i and j are the indices of the current grid; This is the normal vector along the z-axis of the current mesh.

[0029] Step S3.3: Calculate the terrain roughness based on the elevation grid map. Terrain roughness quantifies the degree of irregularity in the terrain surface, i.e., the degree to which the point cloud deviates from the local surface plane. The terrain roughness can be calculated using the following formula:

[0030]

[0031] Among them, z x,y The values ​​are the elevation values ​​of the eight nearest grid cells, with x and y as their indices. The average elevation of the current grid and its neighboring grids;

[0032] Step S3.4: Calculate the terrain ruggedness based on the elevation grid map. As an indicator of the unevenness and complexity of the terrain surface, it can be represented by the average elevation difference between the current grid and its adjacent grids using the following formula:

[0033]

[0034] Among them, z i,j This represents the elevation value of the current grid.

[0035] Step S3.5: Normalize the three terrain indicators obtained in steps S3.2-S3.4 using the min-max method, and obtain the terrain cost by weighting them using the following formula:

[0036] T i,j =w1α′ i,j +w2r′ i,j +w3U′ i,j ;

[0037] Where, α′ i,j 、r′ i,j and U′ i,j These are the normalized results for the terrain parameters slope, roughness, and ruggedness, respectively, with w1, w2, and w3 being the corresponding weighting coefficients.

[0038] Step S3.6: Convert the terrain cost map into a safe-hazard binary image by setting a threshold, and then calculate the safe distance from each safe pixel to the nearest hazard pixel by distance transformation;

[0039] Step S3.7: Based on semantic categories, select areas with a safe distance greater than the physical size of the drone as the final terrain safety area.

[0040] The specific method for step S4 is as follows:

[0041] Step S4.1: Select a linear approach path that meets the preset altitude conditions from the current position of the UAV to any point within the terrain safety area as a candidate path;

[0042] Step S4.2: Discretize the candidate path according to the set sampling interval. Construct a two-dimensional sampling grid along the path direction and perpendicular to the path direction at a fixed interval to form a set of longitudinal and transverse sampling points. Each sampling point represents the spatial risk that the UAV may encounter at that point.

[0043] Step S4.3: Determine whether the cells corresponding to all sampling points in the buffer area determined by the linear approach path collide with obstacles according to the following formula:

[0044] Where, Δz i,j =h i,j -z i,j h represents the vertical distance difference between the j-th lateral sampling point at the i-th vertical position on the path and its corresponding terrain height. i,j z is the flight altitude of the drone at this point. i,j b is the height value of the terrain or obstacle. vertical This is the minimum vertical safety clearance threshold;

[0045] Step S4.4: Further quantitatively evaluate the security of the retained paths using a weighted attenuation cost model, comprehensively considering the security costs of all lateral and longitudinal sampling points along the path, i.e.:

[0046]

[0047] Where N is the number of longitudinal sampling points for the linear approach path, and M is the number of lateral sampling points corresponding to each longitudinal path point; Δz i,j This represents the difference in lateral distance between the sampling point and the obstacle. The exponential function is used to assign a higher risk cost to path points that are closer to obstacles.

[0048] Step S4.5: Path length cost is calculated as the Euclidean distance d between the path origin and the target landing point.e This indicates the energy consumption of a flight path.

[0049] Step S4.6: Combining the path safety cost and the path length cost, construct the following approach cost function as a metric for path quality:

[0050] C=αC a +βd e ;

[0051] Where α and β are weighting coefficients used to balance the safety of the path and the flight cost;

[0052] Step S4.7: The approach cost of all candidate paths is normalized using the min-max normalization method and then weighted and fused with the terrain safety distance to select the optimal landing point that takes into account both terrain safety assessment and approach cost analysis.

[0053] This invention has the following beneficial effects: In an unknown environment, the UAV first uses deep learning to perform semantic segmentation on the RGB images acquired by the visual sensor to obtain semantic images. Then, it fuses the point cloud data collected by the LiDAR with the semantic images to obtain a three-dimensional semantic point cloud, which effectively improves the accuracy and robustness of terrain understanding. Based on semantic labels, it selects terrain safety areas that meet the constraints of terrain geometric features. Finally, by traversing and calculating the cost of the linear approach path of each candidate landing point within the terrain safety area, it selects the landing point with the optimal comprehensive cost, thereby improving the accuracy and practicality of terrain safety area identification. This method not only evaluates the terrain safety of the landing point itself, but also comprehensively considers the path cost of the UAV flying from the current position to the target point, avoiding unreachable or high-risk areas and improving the global point selection optimization effect. Attached Figure Description

[0054] Figure 1 This is a system structure diagram of an embodiment of the present invention;

[0055] Figure 2 This is a flowchart of the method for obtaining semantic point clouds based on semantic images and raw point cloud data according to an embodiment of the present invention;

[0056] Figure 3 This is a flowchart of the terrain-based safe landing area identification method based on semantic point clouds according to an embodiment of the present invention;

[0057] Figure 4 This is a safety-hazard diagram based on terrain cost map conversion as described in an embodiment of the present invention;

[0058] Figure 5 This is a schematic diagram showing the safe distance and optimal landing point of the terrain described in the embodiments of the present invention;

[0059] Figure 6This is a line graph showing the vertical spacing along the approach path direction as described in an embodiment of the present invention. Detailed Implementation

[0060] To clearly illustrate the technical features of the present invention, the present invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings.

[0061] This embodiment presents an autonomous landing method for unmanned aerial vehicles (UAVs) based on terrain assessment and approach cost analysis, enabling safe autonomous landing of UAVs in unknown environments. When the UAV is hovering, it first performs semantic segmentation on RGB images captured by a visual sensor using deep learning to obtain a semantic image. Then, it fuses point cloud data from a LiDAR radar with the semantic image to obtain a 3D semantic point cloud. Based on semantic labels, it filters out safe terrain regions that meet the geometric constraints of the terrain. Next, it iterates through and calculates the approach cost of linear paths to all candidate landing points within the safe terrain region. By comprehensively considering the safe distance of the terrain and the cost of the approach path, the optimal landing point is selected. Finally, it precisely guides the UAV to autonomously land at the target landing point. The system structure of this embodiment is as follows: Figure 1 As shown, the specific steps include:

[0062] Step S1: Real-time acquisition of RGB images via visual sensors and transmission to the onboard computer; semantic segmentation of the acquired RGB images using deep learning; and different colors assigned to various categories for differentiation to obtain semantic images.

[0063] Visual sensors can capture high-resolution image information. Deep learning-based analysis of color and texture patterns helps to identify and classify different target types in ground scenes. This invention uses the following method to perform semantic segmentation on images:

[0064] Step S1.1: Training the semantic segmentation model;

[0065] This embodiment uses the Semantic Drone Dataset as the model training dataset. To address the complexity of the data distribution, various data augmentation strategies, such as image rotation, scaling, flipping, and color perturbation, are employed to enhance the diversity of the training data, thereby improving the generalization ability of the semantic segmentation model. During training, the model classifies pixels in drone aerial images into six semantic categories: roads, buildings, cars, trees, grass, and water. Each category is assigned a specific color label for visual differentiation. The trained semantic segmentation network possesses refined pixel-level feature representation capabilities and excellent semantic discrimination performance, significantly improving segmentation accuracy and robustness.

[0066] Step S1.2: Model deployment and image processing;

[0067] The trained semantic segmentation model is deployed to the UAV's onboard computer, running as the core onboard perception module. In actual missions, the onboard visual sensors acquire RGB images in real time and transmit them to the onboard computer via a high-speed data interface (such as USB or MIPI). The platform then calls the deployed semantic segmentation model to process the input images, performing pixel-level semantic parsing and outputting image results with semantic labels for subsequent analysis and decision-making.

[0068] Step S2: Use the information obtained by the IMU to preprocess the point cloud data collected by the lidar, and combine it with semantic images to achieve information fusion to obtain 3D point cloud data with semantic labels.

[0069] During drone hovering, slight swaying of the flight platform and measurement errors of the sensors themselves can cause noise, distortion, or spatial drift in the point cloud data collected by the lidar, severely affecting the point cloud quality and consequently the accuracy of semantic fusion. Furthermore, raw, dense point clouds increase computational complexity, impacting real-time performance and system stability. Therefore, preprocessing of the point cloud data is necessary before semantic fusion, such as... Figure 2 As shown, the specific steps include:

[0070] Step S2.1: Use software time synchronization to align the vision sensor, IMU and LiDAR using timestamps or ROS time.

[0071] Airborne sensors have different data acquisition frequencies, so it is necessary to time-align the acquired data to ensure that the data are all located near the same point in time. Secondly, due to the differences in the installation position and attitude of visual sensors and lidar, their coordinate systems have spatial deviations. A precise spatial transformation matrix can achieve coordinate unification and ensure accurate spatial mapping of images and point cloud data.

[0072] Multi-sensor time unification; specifically including:

[0073] A time synchronization mechanism based on the ROS platform is adopted, and the ApproximateTimeSynchronizer module in the message_filters package is used to perform software-level alignment of timestamps from multiple sensor topics.

[0074] Calibration of visual sensors and LiDAR; specifically including:

[0075] The external parameters of the vision sensor and LiDAR are calibrated using calibration tools to obtain the transformation relationship of the coordinate system and establish a unified coordinate transformation matrix. The point cloud information of the LiDAR is first projected into the coordinate system of the vision sensor, and then projected into the image coordinate system using the following formula to achieve spatial alignment of multi-source data.

[0076]

[0077] Among them, [x c ,y c ,z c ] T Let R be the 3D position of the point in the visual sensor coordinate system, R be the rotation matrix of the LiDAR coordinate system relative to the visual sensor coordinate system, and t be the translation vector of the LiDAR coordinate system relative to the visual sensor coordinate system. The values ​​of R and t depend on the relative mounting positions of the LiDAR and the visual sensor. [x1, y1, z1] T K represents the 3D position in the point cloud lidar coordinate system, and K is the intrinsic parameter matrix of the camera.

[0078] This example uses the Calibration Toolbox module in Matlab to achieve spatial alignment between multiple sensors in order to obtain the extrinsic parameter matrix of the LiDAR coordinate system relative to the visual sensor coordinate system.

[0079] Step S2.2: Based on the high-frequency attitude information provided by the IMU, perform distortion correction processing on the lidar;

[0080] Within the time window when the lidar completes the acquisition of one frame of point cloud data, the attitude data of the IMU at the corresponding moment is obtained by interpolation. Combined with the emission timestamps of each laser beam inside the lidar, time registration is performed on each point cloud point. Then, based on the motion trajectory and attitude changes of the machine, the real spatial position of each point cloud point at the moment of acquisition is inferred, thereby obtaining point cloud data with motion distortion eliminated, generating high-quality point cloud data with higher spatiotemporal consistency, and providing accurate input for subsequent semantic fusion and terrain analysis.

[0081] Step 2.3: Filter the point cloud data to reduce computational complexity and save computing resources of the airborne computer;

[0082] First, a voxel-based downsampling method is used to perform preliminary sparsification on the point cloud data. The three-dimensional space is divided into several voxel units of predefined size, and the neighborhood of the point cloud within each voxel is aggregated. The geometric centroid of each voxel is calculated as the representative point of that voxel, which significantly reduces the number of points while preserving the shape and boundary features of the point cloud. Then, a statistical filtering algorithm is used to remove outliers from the point cloud data. Based on the distance distribution between each point in the point cloud and its several nearest neighbors, outliers are identified and removed, thereby effectively removing outliers caused by environmental interference, sensor errors, or dynamic factors.

[0083] Step S2.4: Based on the extrinsic matrix of the visual sensor and the lidar, the semantic information in the semantic image is fused into the point cloud information to obtain the semantic point cloud;

[0084] Using this extrinsic parameter matrix, each point in the lidar point cloud coordinate system is projected onto the image plane to obtain its corresponding position in the image coordinate system. Then, the nearest neighbor search algorithm is used to search for the nearest image pixel to each point cloud in the image coordinate system. The semantic information of the pixel is obtained using the semantic image. This semantic information is the semantic information of the corresponding projection point. Finally, the projection point with semantic information is transformed back into the lidar coordinate system to obtain semantic point cloud information, realizing the fusion of image information and point cloud information, and providing a basis for subsequent landing point decision-making.

[0085] Step S3: Perform geometric feature extraction and semantic classification fusion analysis on the semantic point cloud data, calculate the terrain safety distance, and identify suitable terrain safety areas as landing candidate areas, such as... Figure 3 As shown;

[0086] Step S3.1: The semantic point cloud is meshed according to a predetermined resolution. Based on GridMap, the 3D semantic point cloud is used to generate an elevation grid map and a semantic grid map with a resolution of 0.1m / cell to achieve a high-precision discrete representation of the terrain space. Each cell in the grid contains a height value or semantic information. In order to comprehensively evaluate the landability of the terrain, slope, roughness and terrain ruggedness are selected as indicators for terrain evaluation.

[0087] Step S3.2: Calculate the slope of the terrain based on the elevation grid map. Slope is the angle of inclination of the ground relative to the horizontal plane, and is an important indicator of the steepness of the terrain. Based on the elevation grid map, principal component analysis is used to calculate the surface normal of each grid and its neighboring grids in the z-axis direction. Then, the slope of the grid relative to the horizontal plane is obtained by the following formula:

[0088]

[0089] Where i and j are the indices of the current grid; This is the normal vector along the z-axis of the current mesh.

[0090] Step S3.3: Calculate the terrain roughness based on the elevation grid map. Terrain roughness quantifies the degree of irregularity in the terrain surface, i.e., the degree to which the point cloud deviates from the local surface plane. The terrain roughness can be calculated using the following formula:

[0091]

[0092] Among them, z x,y The elevation values ​​of the eight neighboring grids are given by x and y, along with their indices. The average elevation of the current grid and its neighboring grids;

[0093] Step S3.4: Calculate the terrain ruggedness based on the elevation grid map. Terrain ruggedness, as an indicator of the unevenness and complexity of the terrain surface, reflects the severity of ground undulations and changes. It can be represented by the average elevation difference between the current grid and its adjacent grids, calculated using the following formula:

[0094]

[0095] Among them, z i,j This represents the elevation value of the current grid.

[0096] Step S3.5: Normalize the three terrain indicators using the min-max method, and obtain the terrain cost map by weighting them using the following formula:

[0097] T i,j =w1α′ i,j +w2r′ i,j +w3U′ i,j ;

[0098] Where, α′ i,j 、r′ i,j and U′ i,j These are the normalized results for the terrain parameters slope, roughness, and ruggedness, respectively, with w1, w2, and w3 being the corresponding weighting coefficients.

[0099] Step S3.6: Convert the terrain cost map into a safe-hazard binary image by setting a threshold, such as... Figure 4 As shown, areas exceeding a threshold are considered dangerous areas, while those below the threshold are considered safe areas. Then, the distance from each safe pixel to the nearest dangerous pixel, i.e., the safe distance, is calculated using a distance transformation. Figure 5 As shown;

[0100] Step S3.7: Finally, based on semantic categories, select areas with a safe distance greater than the physical size of the drone as the final terrain safety area;

[0101] Step S4: Select the current flight position of the UAV as the starting point, and calculate the cost of the linear approach path of each candidate landing point within the terrain safety area. Then, taking into account the safe distance of the terrain and the cost of the approach path, evaluate each candidate landing point in a weighted manner and select the optimal landing point.

[0102] Step S4.1: First, select a linear approach path that meets the preset altitude conditions from the current position of the UAV to any point within the terrain safety area as a candidate path;

[0103] Step S4.2: Discretize the candidate paths according to the set sampling interval, and introduce the following concepts of lateral buffer (side obstacle avoidance) and vertical buffer (high safety):

[0104] The candidate paths are discretized according to the set sampling interval. A two-dimensional sampling grid is constructed along the path direction and perpendicular to the path direction at a fixed interval, forming a set of longitudinal and transverse sampling points. Each sampling point represents the spatial risk that the UAV may encounter at that point.

[0105] Step S4.3: Path Feasibility Judgment Rules:

[0106] To ensure that the UAV continuously meets the minimum vertical obstacle avoidance requirements during the approach, the following formula is used to determine whether the cells corresponding to all sampling points in the buffer area determined by the linear approach path collide with obstacles:

[0107]

[0108] Where, Δz ij =h ij -z ij h represents the vertical distance difference between the j-th lateral sampling point at the i-th vertical position on the path and its corresponding terrain height. ij z is the flight altitude of the drone at this point. ij b is the height value of the terrain or obstacle. vertical This is the minimum vertical safety clearance threshold;

[0109] like Figure 6 As shown, this embodiment extracts the minimum terrain or obstacle height value from the lateral sampling set at each longitudinal sampling position to construct a sequence of minimum height values ​​arranged along the approach path direction, forming a one-dimensional terrain minimum height data column. This data is then used for point-by-point comparison with the UAV's flight altitude to determine whether the path meets the minimum vertical safety distance requirement. If the vertical safety distance is greater than the threshold, the approach path is deemed feasible and retained; otherwise, the path is considered risky and discarded.

[0110] Step S4.4: Path safety assessment and length cost modeling:

[0111] The retained path undergoes further security quantification assessment. This embodiment employs the following weighted attenuation cost model, comprehensively considering the security costs of all lateral and vertical sampling points along the path, namely:

[0112]

[0113] Where N is the number of longitudinal sampling points for the linear approach path, and M is the number of lateral sampling points corresponding to each longitudinal path point; Δzi,j This represents the difference in lateral distance between the sampling point and the obstacle. The exponential function is used to assign a higher risk cost to path points that are closer to obstacles.

[0114] Step S4.5: Path length cost is calculated as the Euclidean distance d between the path origin and the target landing point. e This indicates the energy consumption of a flight path.

[0115] Step S4.6: Combining the obstacle avoidance cost and the path length cost, construct the following approach cost function as a metric for path quality:

[0116] C=αc a +βd e ;

[0117] Where α and β are weighting coefficients used to balance the safety of the path and the flight cost;

[0118] Step S4.7: Finally, the min-max normalization method is used to normalize the approach costs of all candidate paths, and the normalized result is fused with the terrain safety distance cost to select the optimal landing point that balances terrain safety assessment and approach cost analysis, providing a globally optimal solution for UAV autonomous landing. Figure 5 The marked point in the map is the optimal landing point;

[0119] Step S5: The location information of the selected optimal landing point is used as the target input through coordinate transformation and sent to the flight controller. Based on the target pose and the current state of the UAV, the flight controller adjusts the flight attitude of the UAV in real time, thereby accurately guiding the UAV to autonomously land at the target landing point.

[0120] This example uses the MAVROS framework as an intermediate communication bridge. Based on the MAVLink protocol, it realizes the format conversion of navigation information and the transmission of position commands. The optimal landing point in the ROS world coordinate system is converted into the NED (North-East-Down) coordinate system or GPS global coordinate system (WGS84) standard under the MAVLink protocol through the coordinate mapping interface provided by MAVROS. The flight controller then adjusts the UAV's flight attitude according to the received position information to realize the autonomous landing of the UAV.

[0121] There are many specific ways to implement this invention. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A method for autonomous location selection of unmanned aerial vehicles (UAVs) based on terrain assessment and approach cost analysis, characterized in that, The drone is equipped with onboard visual sensors, an IMU, a lidar, an onboard computer, and a flight controller. After receiving a landing command, the drone autonomously selects its landing location by following these steps: S1: RGB images are acquired in real time through a visual sensor and transmitted to an onboard computer. The acquired RGB images are semantically segmented through deep learning, and different colors are assigned to different categories to distinguish them, thus obtaining semantic images. S2: Use the information obtained by the IMU to preprocess the point cloud data collected by the lidar, and combine it with the semantic image in step S1 to achieve information fusion and obtain three-dimensional point cloud data with semantic labels, i.e. semantic point cloud data. S3: Perform geometric feature extraction and semantic classification fusion analysis on semantic point cloud data, calculate terrain safety distance, and identify suitable terrain safety areas for landing based on semantic labels; S4: Iterate through and calculate the cost of the linear approach path for each candidate landing point within the terrain safety area. Taking into account both the terrain safety distance and the cost of the approach path, select the best landing point through a weighted average. S5: Sends the location information of the selected optimal landing point to the flight controller. The flight controller then precisely guides the drone to autonomously land at the target landing point by adjusting the drone's flight attitude in real time.

2. The UAV autonomous location selection method based on terrain assessment and approach cost analysis according to claim 1, characterized in that, The specific method for step S1 is as follows: Step S1.1: Training the semantic segmentation model; Aerial images were selected as the training dataset. Data augmentation strategies were used to improve the generalization ability of the semantic segmentation model. Image pixels were divided into six semantic categories: roads, buildings, cars, trees, grass, and water, and assigned color labels. Finally, a semantic segmentation model with high segmentation accuracy and strong semantic recognition ability was trained. Step S1.2: Deployment of the semantic segmentation model and image processing; The trained semantic segmentation model is deployed to an onboard computer. RGB images acquired by the visual sensor are transmitted to the onboard computer via an interface. The semantic segmentation model is then used to segment the RGB images and output images with semantic labels.

3. The UAV autonomous location selection method based on terrain assessment and approach cost analysis according to claim 2, characterized in that, The specific method for step S2 is as follows: Step S2.1: Use software time synchronization to align the vision sensor, IMU and LiDAR using timestamps or ROS time. The external parameters of the vision sensor and LiDAR are calibrated using calibration tools to obtain the external parameter matrix. The point cloud information of the LiDAR is first projected into the coordinate system of the vision sensor, and then projected into the image coordinate system to achieve spatial alignment of multi-source data. Among them, [x c, y c ,z c ] T Let R be the 3D position of the point in the visual sensor coordinate system, R be the rotation matrix of the LiDAR coordinate system relative to the visual sensor coordinate system, and t be the translation vector of the LiDAR coordinate system relative to the visual sensor coordinate system. The values ​​of R and t depend on the relative mounting positions of the LiDAR and the visual sensor. [x1, y1, z1] T K represents the 3D position in the point cloud lidar coordinate system, and K is the intrinsic parameter matrix of the camera. Step S2.2: Based on the high-frequency attitude information provided by the IMU, perform distortion correction processing on the lidar; Step S2.3: Filter the point cloud data to reduce computational complexity and save computing resources of the airborne computer; Step S2.4: Based on the extrinsic matrix of the visual sensor and the lidar, the semantic information in the semantic image is fused into the point cloud information to obtain the semantic point cloud; The point cloud is projected onto the image plane using an extrinsic parameter matrix. The semantic information of the corresponding pixels is obtained by nearest neighbor search, and this information is mapped back to the point cloud to generate a semantic point cloud. This achieves the fusion of the image and the point cloud and obtains semantic point cloud data.

4. The UAV autonomous location selection method based on terrain assessment and approach cost analysis according to claim 3, characterized in that, The specific method for step S3 is as follows: Step S3.1: The semantic point cloud is meshed according to a predetermined resolution, and the three-dimensional semantic point cloud is converted into a two-dimensional elevation grid map and semantic grid map, wherein each cell in the grid contains a height value or semantic information. Step S3.2: Calculate the terrain slope based on the elevation grid map. Use principal component analysis to calculate the surface normal of each grid and its neighboring grids in the z-axis direction. Then, obtain the slope of the grid relative to the horizontal plane using the following formula: Where i and j are the indices of the current grid; This is the normal vector along the z-axis of the current mesh. Step S3.3: Calculate the terrain roughness based on the elevation grid map. Terrain roughness quantifies the degree of irregularity in the terrain surface, i.e., the degree to which the point cloud deviates from the local surface plane. The terrain roughness can be calculated using the following formula: Among them, z x,y The values ​​are the elevation values ​​of the eight nearest grid cells, with x and y as their indices. The average elevation of the current grid and its neighboring grids; Step S3.4: Calculate the terrain ruggedness based on the elevation grid map. As an indicator of the unevenness and complexity of the terrain surface, it can be represented by the average elevation difference between the current grid and its adjacent grids using the following formula: Among them, z i,j This represents the elevation value of the current grid. Step S3.5: Normalize the three terrain indicators obtained in steps S3.2-S3.4 using the min-max method, and obtain the terrain cost by weighting them using the following formula: T i,j =w1α′ i,j +w2r′ i,j +w3U′ i,j ; Where, α′ i,j 、r′ i,j and U′ i,j These are the normalized results for the terrain parameters slope, roughness, and ruggedness, respectively, with w1, w2, and w3 being the corresponding weighting coefficients. Step S3.6: Convert the terrain cost map into a safe-hazard binary image by setting a threshold, and then calculate the safe distance from each safe pixel to the nearest hazard pixel by distance transformation; Step S3.7: Based on semantic categories, select areas with a safe distance greater than the physical size of the drone as the final terrain safety area.

5. The UAV autonomous location selection method based on terrain assessment and approach cost analysis according to claim 4, characterized in that: The specific method for step S4 is as follows: Step S4.1: Select a linear approach path that meets the preset altitude conditions from the current position of the UAV to any point within the terrain safety area as a candidate path; Step S4.2: Discretize the candidate path according to the set sampling interval. Construct a two-dimensional sampling grid along the path direction and perpendicular to the path direction at a fixed interval to form a set of longitudinal and transverse sampling points. Each sampling point represents the spatial risk that the UAV may encounter at that point. Step S4.3: Determine whether the cells corresponding to all sampling points in the buffer area determined by the linear approach path collide with obstacles according to the following formula: Where, Δz i,j =h i,j -z i,j h represents the vertical distance difference between the j-th lateral sampling point at the i-th vertical position on the path and its corresponding terrain height. i,j z is the flight altitude of the drone at that point. i,j b is the height value of the terrain or obstacle. vertical This is the minimum vertical safety clearance threshold; Step S4.4: Further quantitatively evaluate the security of the retained paths using a weighted attenuation cost model, comprehensively considering the security costs of all lateral and longitudinal sampling points along the path, i.e.: Where N is the number of longitudinal sampling points for the linear approach path, and M is the number of lateral sampling points corresponding to each longitudinal path point; Δz i,j This represents the difference in lateral distance between the sampling point and the obstacle. The exponential function is used to assign a higher risk cost to path points that are closer to obstacles. Step S4.5: Path length cost is calculated as the Euclidean distance d between the path origin and the target landing point. e This indicates the energy consumption of a flight path. Step S4.6: Combining the path safety cost and the path length cost, construct the following approach cost function as a metric for path quality: C=αc a +βd e 4 Where α and β are weighting coefficients used to balance the safety of the path and the flight cost; Step S4.7: The approach cost of all candidate paths is normalized using the min-max normalization method and then weighted and fused with the terrain safety distance to select the optimal landing point that takes into account both terrain safety assessment and approach cost analysis.

Images