Path planning method and system based on unmanned aerial vehicle surveying and mapping

Abstract

The invention discloses a path planning method and system based on unmanned aerial vehicle surveying and mapping, and relates to the technical field of unmanned aerial vehicle surveying and mapping, and the system comprises a three-dimensional construction unit, a path planning unit and a dynamic obstacle avoidance unit. According to the method, the terrain complexity is scientifically quantified and graded, a quantification basis is provided for subsequent planning, a dividing unit uses a K-means clustering algorithm to accurately divide a surveying and mapping area according to the terrain complexity, the surveying and mapping area is abstracted into a graph structure, a clustering center is used as a node, the weight of an edge is determined in combination with the distance and the terrain complexity, and the surveying and mapping accuracy is improved. The shortest path is searched by means of an A * algorithm, so that the path planning efficiency is effectively improved; the prediction unit optimizes the flight path in the future 30 seconds in a rolling manner with 500 ms as a period, and adjusts the path in real time according to the topographic change, the position of the unmanned aerial vehicle and the obstacle condition, thereby being capable of adapting to terrains with different complexity degrees, reducing the invalid flight distance of the unmanned aerial vehicle, remarkably improving the surveying and mapping efficiency and quality, and reducing the surveying and mapping cost.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) surveying and mapping, and particularly to a path planning method and system based on UAV surveying and mapping. Background Art

[0002] Surveying and mapping engineering, as a key field for obtaining geographical information and spatial data, plays an important role in achieving high-speed development and urban construction. With the development of UAV technology, the application of UAV aerial surveying technology in the surveying and mapping process has become increasingly widespread. UAV aerial surveying technology refers to a set of surveying and mapping technologies that use UAVs as platforms, carry corresponding aerial surveying equipment and sensors, and obtain geographical data and images through aerial photogrammetry technology. It plays an important role in the surveying and mapping process and provides an efficient, flexible, and cost-effective solution for surveying and mapping engineering.

[0003] Currently, in the prior art, in the face of complex and changeable terrains, traditional path planning methods are difficult to efficiently perform reasonable area division and path planning according to terrain complexity, and problems such as overly long paths, serious detours, or inability to adapt to terrain changes are likely to occur, resulting in low UAV surveying and mapping efficiency.

[0004] Therefore, a path planning method and system based on UAV surveying and mapping are proposed to solve the above problems. Summary of the Invention

[0005] The main purpose of the present invention is to provide a path planning method and system based on UAV surveying and mapping to solve the problems raised in the above background.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is: a path planning method and system based on UAV surveying and mapping, the system includes a three-dimensional construction unit, a path planning unit, and a dynamic obstacle avoidance unit; The three-dimensional construction unit is used to collect multi-modal data by a collection device, generate final point cloud data after preprocessing, and calculate multi-dimensional terrain feature data based on it; The path planning unit determines terrain complexity based on multi-dimensional terrain feature data, divides the surveying and mapping area through the K-means clustering algorithm based on terrain complexity, generates an initial path in combination with the graph search algorithm, and simultaneously optimizes the future 30-second flight track in a rolling manner with a period of 500 ms; The dynamic obstacle avoidance unit constructs a probability occupancy grid map through millimeter-wave radar and visual data, predicts the movement trajectory of obstacles using a deep learning model, and optimizes the obstacle avoidance path using a third-order Bezier curve.

[0007] Preferably, the three-dimensional construction unit includes a collection unit, a preprocessing unit, a generation unit, and a feature unit; The acquisition unit is used to obtain the distance information of the terrain through lidar measurement, the binocular vision sensor to obtain stereo vision information through the parallax principle, the IMU device to measure the attitude information of the UAV, and the RGB-D camera to obtain the color image and depth information of the terrain; The preprocessing unit is used to preprocess the multi-modal data collected by the acquisition unit. The data collected by the RGB-D camera and the IMU device are converted into point cloud data through parallax calculation. The data collected by the lidar is directly used to obtain three-dimensional point cloud data. The data collected by the IMU device is used to calibrate the sensor attitude, including error compensation for translation and rotation; The generation unit is used to unify the coordinate systems of the acquisition devices through the hand-eye calibration method, and weight and fuse the point cloud data of the acquisition devices according to the coordinates to generate the final point cloud data; The feature unit calculates multi-dimensional terrain feature data based on the final point cloud data generated by weighted fusion, including calculating slope data based on the normal estimation algorithm, roughness calculated by the local standard deviation algorithm, and curvature approximately calculated by the second-order difference of the heights of neighboring points.

[0008] Preferably, the path planning unit includes a complexity unit, a division unit, an initial unit, and a prediction unit; The complexity unit calculates the comprehensive terrain complexity value through weighted summation based on the multi-dimensional terrain feature data of slope, roughness, and curvature, and divides different complexity levels according to the threshold ranges of each index; The division unit divides the surveying and mapping area through the K-means clustering algorithm. The steps are as follows: Step 1: Initially set an appropriate K value according to the terrain complexity, randomly select K points from the final point cloud data of the surveying and mapping area as the initial cluster centers, and calculate the distances from each data point in the surveying and mapping area to the K cluster centers; Step 2: Assign each data point to the cluster where the nearest cluster center is located. For each cluster, calculate the average value of the coordinates of all data points within the cluster to obtain the new cluster center position; Step 3: Calculate the distance between the new cluster centers and the previous round of cluster centers through the Euclidean distance method. If the moving distances of all cluster centers are less than the pre-set threshold, it is considered that the clustering converges and the algorithm terminates. Otherwise, return to Step 2 to continue iterative calculation; Step 4: After the algorithm terminates, different clusters are the different surveying and mapping areas that have been divided.

[0009] Preferably, the initial unit abstracts the survey area into a graph, where the nodes are set as the clustering centers of key position points, the edges represent the connection relationships between the nodes, and the weights of the edges are determined according to the distance and terrain complexity. The defined nodes, edges, and weights are stored using an adjacency matrix to form a complete graph structure. The A* algorithm is used to set the starting point and the ending point, and the Manhattan distance and cost function are used to search for the shortest path from the starting point to the ending point; The prediction unit acquires the current position, attitude, and the latest terrain and obstacles every 500 ms. Based on this, it re-evaluates the current path cost, adjusts the path using a local replanning algorithm, and generates a global path.

[0010] Preferably, the dynamic obstacle avoidance unit includes a grid unit, a trajectory unit, an avoidance unit, and an adjustment unit.

[0011] Preferably, the grid unit divides the environment around the UAV on the global path into uniformly sized grid cells. Each grid corresponds to a small area in the actual environment. The lidar data is converted into a point cloud form and projected onto the grid map. The corresponding grid is determined according to the position of the points. At the same time, the data collected by the binocular vision sensor detects obstacles in the image through YOLO and maps the detected obstacle bounding boxes onto the grid map; the Bayesian method is used to fuse the two types of data to update the occupancy probability of each grid.

[0012] Preferably, the trajectory unit uses a spatio-temporal convolutional network, inputs the data collected by the lidar and the binocular vision sensor for 5 consecutive frames, and outputs the trajectory prediction of the obstacle in the next 2 seconds.

[0013] Preferably, the avoidance unit determines the current position of the UAV as the starting point and the target position as the ending point based on the probability occupancy grid map, and uses the A* algorithm to generate a preliminary obstacle avoidance path. Several control points are selected on the preliminary obstacle avoidance path, including the starting point, the ending point, and the path turning points. The third-order Bezier curve formula is used to generate a smooth curve to replace the sharp turning parts in the original path to form the final path.

[0014] Preferably, the adjustment unit evaluates the optimized path according to the length of the final path, the safe distance from the obstacles, and the UAV motion constraints including the maximum turning radius and the maximum speed. If the requirements are not met, the positions of the control points are adjusted, and the Bezier curve is regenerated until the optimal obstacle avoidance path is obtained. The optimal obstacle avoidance path is transmitted to the prediction unit to update the global path based on the optimal obstacle avoidance path, realizing path planning.

[0015] A path planning method based on UAV surveying and mapping includes the following steps: Step 1: Process the multi-modal data to generate the final point cloud data and calculate the multi-dimensional terrain feature data; Step 2: Divide the surveying and mapping area based on terrain complexity, generate an initial planned path, and simultaneously optimize the path iteratively. Step 3: Construct a grid map to predict obstacles, optimize the path with curves to obtain the optimal obstacle avoidance path, and update the global path based on the optimal obstacle avoidance path.

[0016] The present invention has the following beneficial effects: 1. In the present invention, the acquisition unit uses various devices such as lidar and binocular vision sensors to obtain terrain information from multiple dimensions including distance, stereo vision, attitude, color image, and depth, realizing a comprehensive perception of the surveying and mapping environment. The preprocessing unit preprocesses data according to the characteristics of different devices, unifies the coordinate system through hand-eye calibration, and weighted-fuses the point cloud data to avoid data conflicts. The feature unit, based on accurate point cloud data, uses professional algorithms to calculate terrain features such as slope, roughness, and curvature, providing high-precision data support for subsequent path planning, ensuring that the planned path fits the actual terrain, and improving the reliability and accuracy of path planning.

[0017] 2. In the present invention, the complexity unit scientifically quantifies the terrain complexity and divides it into levels by weighted summing of multi-dimensional terrain feature data such as slope, roughness, and curvature, providing a quantitative basis for subsequent planning; the division unit uses the K-means clustering algorithm to accurately divide the surveying and mapping area according to the terrain complexity, abstracts the surveying and mapping area into a graph structure, determines the weight of the edge with the clustering center as the node, combined with distance and terrain complexity, and then uses the A* algorithm to search for the shortest path, effectively improving the path planning efficiency; the prediction unit iteratively optimizes the future 30-second flight path with a period of 500 ms, adjusts the path in real time according to terrain changes, UAV position, and obstacle conditions, so as to be able to adapt to terrains of different complexity levels, reduce the ineffective flight distance of the UAV, significantly improve the surveying and mapping efficiency and quality, and reduce the surveying and mapping cost.

[0018] 3. In the present invention, the grid unit divides the environment around the UAV into grid cells based on the global path, fuses lidar and binocular vision data, uses the Bayesian method to update the grid occupancy probability, and constructs a probability occupancy grid map that accurately reflects the obstacle distribution. The trajectory unit uses a spatio-temporal convolutional network to input continuous multi-frame sensor data to accurately predict the future motion trajectory of the obstacle, and generates a preliminary obstacle avoidance path based on the probability occupancy grid map, and then optimizes it with a third-order Bezier curve to make the path smooth and conform to the motion characteristics of the UAV. The adjustment unit evaluates the path according to the path length, safety distance, and UAV motion constraints, and repeatedly optimizes it until the optimal obstacle avoidance path is obtained and the global path is updated, enabling the UAV to timely respond to dynamic obstacles, effectively reducing the collision risk, and improving the safety and reliability of UAV operations. Description of the Drawings

[0019] Figure 1It is a flowchart of a path planning method based on UAV mapping according to the present invention; Figure 2 It is a flowchart of a path planning system based on UAV mapping according to the present invention. Specific embodiments

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0021] Please refer to Figure 1 and Figure 2 , the present invention provides a technical solution: a path planning method and system based on UAV mapping. The system includes a three-dimensional construction unit, a path planning unit, and a dynamic obstacle avoidance unit; The three-dimensional construction unit is used to collect multi-modal data by a collection device, generate final point cloud data after preprocessing and fusion, and calculate multi-dimensional terrain feature data based on it; The path planning unit judges the terrain complexity based on the multi-dimensional terrain feature data, divides the mapping area through the K-means clustering algorithm based on the terrain complexity, generates an initial path in combination with the graph search algorithm, and simultaneously optimizes the future 30-second flight track in a cycle of 500 ms; The dynamic obstacle avoidance unit constructs a probability occupancy grid map through millimeter-wave radar and visual data, predicts the movement trajectory of obstacles using a deep learning model, and optimizes the obstacle avoidance path using a third-order Bezier curve.

[0022] The three-dimensional construction unit includes a collection unit, a preprocessing unit, a generation unit, and a feature unit; The collection unit is used to measure the distance information of the terrain through lidar, obtain stereo vision information through the binocular vision sensor based on the parallax principle, measure the attitude information of the UAV by the IMU device, and obtain the color image and depth information of the terrain by the RGB-D camera; The preprocessing unit is used to preprocess the multi-modal data collected by the collection unit, including converting the data collected by the RGB-D camera and the IMU device into point cloud data through parallax calculation, directly obtaining three-dimensional point cloud data from the data collected by the lidar, and using the data collected by the IMU device to calibrate the sensor attitude, including error compensation for translation and rotation; The generation unit is used to unify the coordinate systems of the collection devices through the hand-eye calibration method, and weighted-fuse the point cloud data of the collection devices according to the coordinates to generate the final point cloud data; The feature unit calculates multi-dimensional terrain feature data based on the final point cloud data generated by weighted fusion, including calculating slope data based on the normal estimation algorithm, calculating roughness through the local standard deviation algorithm, and calculating curvature by approximately calculating the second-order difference of the heights of neighboring points.

[0023] Specifically, multi-modal geographic information data is collected collaboratively by lidar, binocular vision sensors, IMU devices, and RGB-D cameras, respectively obtaining terrain geometric structures, stereo vision information, attitude parameters, and color depth data. During the preprocessing process, the image data of the RGB-D camera is combined with the attitude data of the IMU, and point clouds with position calibration are generated through disparity calculation; the lidar directly outputs three-dimensional point clouds without conversion, and the IMU data is used to compensate for the rotation and translation errors generated by the movement of the drone, ensuring the spatial alignment of the data of each sensor. The generation unit establishes a unified coordinate system framework through hand-eye calibration, performs coordinate transformation on the point clouds output by different sensors, and then performs fusion processing according to the confidence weights to form a final point cloud with high precision and high integrity. The feature unit is based on the fused point cloud, calculates the slope using normal estimation to reflect the terrain inclination degree, statistically calculates the surface roughness using the local standard deviation, extracts the curvature features through second-order difference approximation, and constructs a multi-dimensional terrain feature dataset to provide a basis for subsequent terrain complexity grading.

[0024] The path planning unit includes a complexity unit, a division unit, an initial unit, and a prediction unit; The complexity unit calculates the comprehensive terrain complexity value through weighted summation based on the multi-dimensional terrain feature data of slope, roughness, and curvature, and divides different complexity levels according to the threshold ranges of each index; The division unit divides the surveying and mapping area through the K-means clustering algorithm. The steps are as follows: Step 1: Initially set an appropriate K value according to the terrain complexity, randomly select K points from the final point cloud data of the surveying and mapping area as the initial clustering centers, and calculate the distances from each data point in the surveying and mapping area to the K clustering centers; Step 2: Assign each data point to the cluster where the nearest clustering center is located. For each cluster, calculate the average value of the coordinates of all data points within the cluster to obtain the new position of the clustering center. Step 3: Calculate the distance between the new clustering centers and the clustering centers in the previous round through the Euclidean distance method. If the moving distances of all clustering centers are less than the pre-set threshold, it is considered that the clustering converges and the algorithm terminates; otherwise, return to Step 2 to continue iterative calculation; Step 4: After the algorithm terminates, the different clusters are the different surveying and mapping areas that have been divided.

[0025] Specifically, the weighted summation calculation of the comprehensive terrain complexity value refers to linearly combining the three terrain feature indicators of slope, roughness, and curvature according to preset weights. Specifically, it can be achieved by normalizing each indicator and then multiplying it by the corresponding weight coefficient and adding them together. For example, the slope weight is set to 0.5, the roughness is 0.3, and the curvature is 0.2. By fusing multi-dimensional indicators, the accuracy of terrain assessment is improved. The K-means clustering algorithm refers to an unsupervised machine learning method based on the characteristics of data distribution. Specifically, it can be achieved by randomly initializing the clustering centers and iteratively optimizing the cluster division, and is used to dynamically divide the mapping area according to the terrain complexity. The Euclidean distance method refers to a mathematical method for calculating the straight-line distance between two points. Specifically, it can be achieved by using the formula of summing the squared differences and then taking the square root, and is used to determine whether the clustering centers tend to a stable state; The complex unit first normalizes the three terrain feature data of slope, roughness, and curvature, eliminates the dimensional difference, and then performs weighted summation according to the preset weights to generate a scalar value reflecting the regional comprehensive complexity. For example, in an area with steep slopes and rough surfaces, the weighted comprehensive value will be significantly higher than that in a flat area. Subsequently, according to the preset complexity level threshold, the mapping area is divided into three complexity types: high, medium, and low. The division unit automatically determines the initial number of clusters based on the complexity distribution characteristics. For example, a larger K value is set in an area with dense complex terrain to achieve fine division. After randomly selecting the initial clustering centers, the Euclidean distance from each data point to each center is calculated, and the point is assigned to the cluster corresponding to the nearest center. After each iteration, the centroid position of the cluster is recalculated until the distance between the new and old centroids is less than the set threshold. At this time, the cluster boundary forms a spatial match with the terrain complexity distribution, and finally, multiple sub-regions with similar terrain features are output.

[0026] The initial unit abstracts the mapping area as a graph, where the nodes are set as the key position point clustering centers, the edges represent the connection relationships between the nodes, and the weights of the edges are determined according to the distance and terrain complexity. The defined nodes, edges, and weights are stored using the adjacency matrix to form a complete graph structure. The A* algorithm is used to set the starting point and the ending point, and the Manhattan distance and the cost function are used to search for the shortest path from the starting point to the ending point; The prediction unit obtains the current position, attitude, and the latest terrain and obstacles every 500ms. Based on this, it re-evaluates the current path cost and uses a local replanning algorithm, such as the Dijkstra algorithm, to adjust the path and generate a global path.

[0027] Specifically, after completing the regional clustering division, the regional clustering centers are used as path nodes to construct a graph structure; the connection relationship between each pair of nodes is determined according to the accessibility of the UAVs. When there is a passable path between the regions where two nodes are located, a connection edge is established; the weight value of the edge is obtained by multiplying the straight-line distance between the nodes by the distance coefficient and adding the mean value of the terrain complexity of the regions where the two nodes are located multiplied by the complexity coefficient; the weight value of the non-connected nodes in the adjacency matrix is set to infinity to form a sparse matrix storage structure; when the A* algorithm is executed, the node with the minimum comprehensive cost is preferentially expanded, where the comprehensive cost is the sum of the actual cost of the traveled path and the estimated remaining cost of the Manhattan distance; when a terrain change or a new obstacle is detected, the prediction unit triggers local replanning, and on the premise of keeping the global path framework unchanged, the local optimal path is recalculated by adjusting the connection relationship and weight value of the adjacent nodes.

[0028] The dynamic obstacle avoidance unit includes a grid unit, a trajectory unit, an avoidance unit, and an adjustment unit.

[0029] The grid unit divides the environment around the UAV on the global path into uniformly sized grid cells. Each grid corresponds to a small area in the actual environment. The lidar data is converted into a point cloud form and projected onto the grid map. The corresponding grid is determined according to the position of the points. At the same time, the data collected by the binocular vision sensor identifies the obstacles in the image through YOLO and maps the detected obstacle bounding boxes onto the grid map; the Bayesian method is used to fuse the two types of data to update the occupancy probability of each grid.

[0030] Specifically, after the environmental space is divided into uniform grids, the lidar point cloud accurately fills the corresponding grids through coordinate projection, directly reflecting the spatial position of the obstacles; the binocular vision data identifies the obstacle bounding boxes through the object detection algorithm and maps them into the grid map to supplement the semantic information. The two types of heterogeneous sensor data are spatially aligned under the unified grid framework, and then the occupancy probability value of each grid is dynamically updated through the Bayesian probability formula. When the lidar detects an obstacle in a certain grid, the occupancy probability of that grid is updated positively; when the vision detection confirms the obstacle category, the probability value is further corrected. Through the complementarity and cooperation of multi-source data, both the geometric accuracy of the lidar and the semantic understanding ability of the vision are retained.

[0031] The trajectory unit uses a spatio-temporal convolutional network to input the data collected by the lidar and the binocular vision sensor for 5 consecutive frames and outputs the trajectory prediction of the obstacles in the next 2 seconds.

[0032] Specifically, the spatio-temporal convolutional network extracts the motion trends in five consecutive frames of data through convolutional operations in the time dimension, and analyzes the shape and position changes of obstacles in the space dimension at the same time; the lidar point cloud data forms three-dimensional grid features after systematic processing, and is fused with the two-dimensional image features extracted by binocular vision through a cross-modal attention mechanism to form a spatio-temporal joint feature map. The trajectory prediction results output by the network include the predicted coordinates and confidence levels of obstacles at each time step within the next 2 seconds. These data are input into the path optimization module as dynamic constraint conditions. During the calculation process, the spatio-temporal convolutional layer maintains time continuity through a sliding window mechanism, and the residual connection structure avoids the problem of gradient disappearance in deep networks.

[0033] Based on the probabilistic occupancy grid map, the avoidance unit determines the current position of the UAV as the starting point and the target position as the ending point, and uses the A* algorithm to generate a preliminary obstacle avoidance path. Several control points are selected on the preliminary obstacle avoidance path, including the starting point, the ending point, and the path turning points. The third-order Bezier curve formula is used to generate a smooth curve to replace the sharp turning parts in the original path to form the final path.

[0034] Among them, the third-order Bezier curve formula is used to generate a smooth curve, and its formula is: ; Among them, and are control points, is the ending point, represents the parameter variable.

[0035] Specifically, this scheme first identifies the distribution of obstacles through the probabilistic occupancy grid map, and uses the A* algorithm to quickly generate a broken-line path containing necessary obstacle avoidance nodes. Subsequently, the starting point, the ending point, and the path turning points are selected from the path as control points, and the discrete path points are mathematically fitted through the third-order Bezier curve equation; this curve converts the straight-line segments between adjacent control points into curve segments with continuously changing curvature through parametric modeling, while retaining the key obstacle avoidance nodes of the original path, eliminating the path discontinuity problem caused by sharp turns; by adjusting the position relationship of the control points, a smooth flight trajectory that meets the maximum turning radius constraint of the UAV can be generated, thus realizing the dynamic matching of the path geometric features and kinematic constraints.

[0036] The adjustment unit evaluates the optimized path according to the final path length, the safe distance from obstacles, and the UAV motion constraints including the maximum turning radius and the maximum speed. If the requirements are not met, the positions of the control points are adjusted, and the Bezier curve is regenerated until the optimal obstacle avoidance path is obtained, and the optimal obstacle avoidance path is transmitted to the prediction unit to update the global path based on the optimal obstacle avoidance path, realizing path planning.

[0037] Specifically, during the path evaluation phase, three core indicators are calculated simultaneously: the total path length, the minimum distance to obstacles, and the compliance of the turning radius with the speed limit. Among them, the path length affects the flight time efficiency, the safety distance determines the reliability of obstacle avoidance, and the motion constraint parameters ensure the executability of the path. When any of the indicators exceeds the preset threshold, the path correction process is triggered. During the path correction process, the curve shape is changed by adjusting the control point coordinates of the Bezier curve. For example, when the turning radius is insufficient, intermediate control points are added to expand the curve curvature radius, and when the safety distance is insufficient, the control points are translated to offset the original path trajectory. After each adjustment, a new Bezier curve is generated and evaluated again until all indicators meet the requirements. The optimized path is transmitted to the prediction unit in real time, and the global flight route is updated by replacing the corresponding section in the original path to ensure that the UAV can respond to environmental changes in a timely manner.

[0038] In the present invention, a path planning system based on UAV mapping, the acquisition unit uses various devices such as lidar and binocular vision sensors to obtain terrain information from multiple dimensions including distance, stereo vision, attitude, color image, and depth, realizing a comprehensive perception of the mapping environment. The preprocessing unit preprocesses the data according to the characteristics of different devices, unifies the coordinate system through hand-eye calibration, and weighted-fuses the point cloud data to avoid data conflicts. The feature unit, based on accurate point cloud data, uses professional algorithms to calculate terrain features such as slope, roughness, and curvature, providing high-precision data support for subsequent path planning, ensuring that the planned path fits the actual terrain, and improving the reliability and accuracy of path planning. The complexity unit scientifically quantifies the terrain complexity and divides it into levels by weighted summing of multi-dimensional terrain feature data such as slope, roughness, and curvature, providing a quantitative basis for subsequent planning; the division unit uses the K-means clustering algorithm to accurately divide the mapping area according to the terrain complexity, abstracts the mapping area into a graph structure, determines the weight of the edge with the clustering center as the node, combined with distance and terrain complexity, and then searches for the shortest path by means of the A* algorithm, effectively improving the path planning efficiency; the prediction unit optimizes the future 30-second flight path in a rolling manner with a period of 500 ms, and adjusts the path in real time according to terrain changes, UAV position, and obstacle conditions, so as to be able to adapt to terrains of different complexities, reduce the ineffective flight distance of the UAV, significantly improve the mapping efficiency and quality, and reduce the mapping cost. Based on the global path, the grid unit divides the environment around the UAV into grid units, fuses lidar and binocular vision data, updates the grid occupancy probability using the Bayesian method, and constructs a probability occupancy grid map that accurately reflects the distribution of obstacles. The trajectory unit uses a spatio-temporal convolutional network to input consecutive multi-frame sensor data to accurately predict the future movement trajectory of the obstacles, and based on the probability occupancy grid map, generates a preliminary obstacle avoidance path, which is then optimized by a third-order Bezier curve to make the path smooth and conform to the movement characteristics of the UAV. The adjustment unit evaluates the path according to the path length, safety distance, and UAV movement constraints, repeatedly optimizes it until the optimal obstacle avoidance path is obtained, and updates the global path, enabling the UAV to respond to dynamic obstacles in a timely manner, effectively reducing the collision risk, and improving the safety and reliability of UAV operations.

[0039] A path planning method based on UAV mapping, comprising the following steps: Step 1: Collect multi-modal data, process it to generate final point cloud data, and calculate multi-dimensional terrain feature data; Step 2: Divide the mapping area based on terrain complexity, generate an initial planned path, and simultaneously optimize the path iteratively; Step 3: Construct a grid map to predict obstacles, optimize the path with a curve to obtain the optimal obstacle avoidance path, and update the global path based on the optimal obstacle avoidance path.

[0040] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A path planning system based on drone mapping, characterized in that, The system includes a 3D construction unit, a path planning unit, and a dynamic obstacle avoidance unit; The 3D construction unit is used for the acquisition device to obtain multimodal data, which is fused to generate final point cloud data after preprocessing, and calculate multi-dimensional terrain feature data based on it; The path planning unit judges the terrain complexity based on the multi-dimensional terrain feature data, divides the survey area through the K-means clustering algorithm based on the terrain complexity, generates an initial path in combination with the graph search algorithm, and simultaneously optimizes the future 30-second flight track in a cycle of 500 ms; The dynamic obstacle avoidance unit constructs a probability occupancy grid map through millimeter-wave radar and visual data, predicts the movement trajectory of obstacles using a deep learning model, and optimizes the obstacle avoidance path using a third-order Bezier curve.

2. The path planning system based on UAV mapping according to claim 1, wherein The 3D construction unit includes an acquisition unit, a preprocessing unit, a generation unit, and a feature unit; The acquisition unit is used to measure the distance information of the terrain through a lidar, obtain stereo vision information through a binocular vision sensor based on the parallax principle, measure the attitude information of the UAV by an IMU device, and obtain the color image and depth information of the terrain by an RGB-D camera; The preprocessing unit is used to preprocess the multimodal data collected by the acquisition unit, including converting the data collected by the RGB-D camera and the IMU device into point cloud data through parallax calculation, directly obtaining 3D point cloud data from the data collected by the lidar, and using the data collected by the IMU device to calibrate the sensor attitude, including error compensation for translation and rotation; The generation unit is used to unify the coordinate systems of the acquisition devices through the hand-eye calibration method, and weighted-fuse the point cloud data of the acquisition devices according to the coordinates to generate final point cloud data; The feature unit calculates multi-dimensional terrain feature data based on the final point cloud data generated by weighted fusion, including calculating slope data based on the normal estimation algorithm, calculating roughness through the local standard deviation algorithm, and calculating curvature by approximately calculating the second-order difference of the heights of neighboring points; 3. The path planning system based on UAV mapping according to claim 1, wherein The path planning unit includes a complexity unit, a division unit, an initial unit, and a prediction unit; The complexity unit calculates the comprehensive terrain complexity value through weighted summation based on the multi-dimensional terrain feature data of slope, roughness, and curvature, and divides different complexity levels according to the threshold ranges of each index; The division unit divides the survey area through the K-means clustering algorithm, and the steps are as follows: Step 1: Initially set an appropriate K value according to the terrain complexity, randomly select K points from the final point cloud data of the survey area as the initial clustering centers, and calculate the distances from each data point in the survey area to the K clustering centers; Step 2: Assign each data point to the cluster where the nearest clustering center is located. For each cluster, calculate the average value of the coordinates of all data points in the cluster to obtain the new clustering center position; Step 3: Calculate the distance between the new clustering center and the previous round of clustering centers through the Euclidean distance method. If the moving distances of all clustering centers are less than a pre-set threshold, it is considered that the clustering converges and the algorithm terminates. Otherwise, return to Step 2 to continue iterative calculation; Step 4: After the algorithm terminates, different clusters are the divided different survey areas.

4. A path planning system based on UAV mapping according to claim 3, characterized in that, The initial unit abstracts the survey area into a graph, where the nodes are set as the clustering centers of key position points, the edges represent the connection relationships between the nodes, and the weights of the edges are determined according to the distance and terrain complexity. The defined nodes, edges, and weights are stored using an adjacency matrix to form a complete graph structure. The A* algorithm is used to set the starting point and the ending point, and the Manhattan distance and cost function are used to search for the shortest path from the starting point to the ending point; The prediction unit obtains the current position, attitude, and the latest terrain and obstacles every 500 ms. Based on this, it re-evaluates the current path cost, adjusts the path using a local replanning algorithm, and generates a global path.

5. A path planning system based on UAV mapping according to claim 1, characterized in that, The dynamic obstacle avoidance unit includes a grid unit, a trajectory unit, an avoidance unit, and an adjustment unit.

6. The path planning system based on UAV mapping according to claim 5, characterized in that, The grid unit divides the environment around the UAV on the global path into uniformly sized grid cells. Each grid corresponds to a small area in the actual environment. The lidar data is converted into a point cloud form and projected onto the grid map. The corresponding grid is determined according to the position of the points. At the same time, the data collected by the binocular vision sensor uses YOLO to identify obstacles in the image, and the detected obstacle bounding boxes are mapped onto the grid map; Bayesian methods are used to fuse the two types of data to update the occupancy probability of each grid.

7. A path planning system based on UAV mapping according to claim 5, characterized in that, The trajectory unit uses a spatio-temporal convolutional network, inputs the data collected by the lidar and the binocular vision sensor for 5 consecutive frames, and outputs the trajectory prediction of the obstacles in the next 2 seconds.

8. A path planning system based on UAV mapping according to claim 5, characterized in that, The avoidance unit, based on the probabilistic occupancy grid map, determines the current position of the UAV as the starting point and the target position as the ending point, and uses the A* algorithm to generate a preliminary obstacle avoidance path. Several control points are selected on the preliminary obstacle avoidance path, including the starting point, the ending point, and the path turning points. The third-order Bezier curve formula is used to generate a smooth curve to replace the sharp turning parts in the original path to form the final path.

9. A path planning system based on UAV mapping according to claim 5, characterized in that, The adjustment unit evaluates the optimized path according to the length of the final path, the safe distance from the obstacles, and the UAV motion constraints including the maximum turning radius and the maximum speed. If the requirements are not met, the positions of the control points are adjusted, and the Bezier curve is regenerated until the optimal obstacle avoidance path is obtained. The optimal obstacle avoidance path is transmitted to the prediction unit to update the global path based on the optimal obstacle avoidance path, thus realizing path planning.

10. A path planning method based on UAV mapping, which refers to the path planning system based on UAV mapping described in any one of claims 1-9, characterized in that, It includes the following steps: Step 1: Process the multi-modal data to generate the final point cloud data and calculate the multi-dimensional terrain feature data; Step 2: Divide the survey area based on the terrain complexity and generate an initial planned path, and at the same time, roll-optimize the path; Step 3: Construct a grid map to predict obstacles, optimize the path with a curve to obtain the optimal obstacle avoidance path, and update the global path based on the optimal obstacle avoidance path.

Images