Unmanned aerial vehicle grid map construction method for multi-view rapid detection of following target

Abstract

The invention belongs to the technical field of unmanned aerial vehicle navigation, and particularly relates to an unmanned aerial vehicle grid map construction method for multi-view rapid detection of a following target, and the method comprises the steps: employing the flight state of a current unmanned aerial vehicle to adaptively adjust the frequency of depth map information in front, left and right directions; the depth map information in the three directions after frequency adjustment is fused to obtain a local point cloud map; based on the local point cloud map, the space is rasterized, the feature occupation probability of each grid in each grid map is calculated, the point cloud is mapped to the corresponding grid, and a static obstacle map is obtained. On the basis, a map dynamic information enhancement strategy is designed, namely, dynamic obstacle recognition and state estimation are achieved through information such as point cloud space distribution and map occupancy probability, and the accuracy of dynamic obstacle construction in a local map is improved. In conclusion, the dynamic obstacle map is accurately constructed, meanwhile, the mapping range is increased, the accuracy is improved, and the unnecessary calculation amount is effectively reduced.

Description

Technical Field

[0001] The invention belongs to the technical field of unmanned aerial vehicle navigation, and is a method for constructing an unmanned aerial vehicle grid map for rapidly detecting and following targets from multiple perspectives. Background Art

[0002] Drones are highly favored in the field of target following due to their small size, high maneuverability, and low cost. In highly dynamic target following scenarios, there are two main factors that affect the rapid and robust following of drones: first, the visual system's perception of the environment is inaccurate and incomplete, especially in dynamic scenarios, making it difficult for drones to build accurate and comprehensive local maps in a timely manner; second, drones are prone to losing their targets during highly dynamic maneuvers, and are unable to correctly build targets in local maps. Therefore, large-scale and accurate map construction is a prerequisite for drones to achieve safe target following.

[0003] Current map construction schemes mainly include maps based on visual features, maps based on distance functions, and maps based on point clouds. However, in maps based on visual features, because the point clouds are too sparse and lack higher-order descriptions of features, the method using deep learning models can not only only recognize pre-trained objects, but also occupy a large amount of onboard computing resources. Maps based on distance functions have a limited range, contribute little to planning, and cannot predict the trajectory of dynamic objects. Maps based on point clouds need to balance computing costs and data accuracy and optimize the original point clouds, but point clouds are very sensitive to sensor inputs, and environmental noise greatly affects the accuracy of mapping. In summary, the existing perception algorithms for autonomous and safe navigation of drones have limitations in terms of expression and the types of objects that can be identified. Summary of the invention

[0004] The purpose of the present invention is to provide a method for constructing a raster map of an unmanned aerial vehicle (UAV) for rapid detection and follow-up of targets from multiple perspectives, so as to overcome the limitations of existing perception algorithms for autonomous and safe navigation of UAVs in terms of expression methods and identifiable object types.

[0005] To achieve the above object, the present invention adopts the following technical solution: A method for constructing a raster map of an unmanned aerial vehicle for rapid detection and tracking of a target from multiple perspectives comprises the following steps: Step 1: Collect the flight status of the drone and the depth map of the surrounding environment in real time. The flight status includes the flight speed, airspeed, and yaw angle. The depth map of the surrounding environment is a three-way depth map perceived by binocular cameras installed in the front, left, and right directions of the drone. Step 2: dynamically adjust the frequency of the three-dimensional depth map information according to the flight status of the UAV, fuse the three-dimensional depth map after the frequency adjustment to obtain a local point cloud map; rasterize the space based on the local point cloud map to obtain a grid local map; use the grid posterior probability algorithm to calculate the probability value of each grid in the grid local map, and use the probability value to update the occupied grid local map to obtain a static obstacle map; Step 3: Cluster the point clouds corresponding to different objects in the local point cloud map to obtain the position of the dynamic object; estimate the actual position of the dynamic object based on the uniform speed assumption of the dynamic obstacle between the two frames of the depth map; predict the grid position of the corresponding dynamic point cloud mapped to the static obstacle map according to the actual position of the dynamic object, and calculate its probability value to update the static obstacle map to obtain the dynamic obstacle map; Step 4: Keep the drone positioning grid at the center grid, fix the direction of the map to the initial direction, update the occupied grid local map according to the current flight status of the drone, and rebuild the grid local map according to steps 1-3 for the new local area after the drone displacement change until the target point is reached to complete navigation and obstacle avoidance.

[0006] Furthermore, the principle of dynamically adjusting the frequency of the three-way depth map information according to the flight status of the drone in step 2 is: ; In the formula, is the frame rate of the left camera, is the frame rate of the camera on the right, The flight speed of the drone, is the flight speed direction of the drone; is the heading angle of the UAV, is the difference between the UAV flight speed direction and the UAV heading angle. The angles in this formula are defined in the spatial coordinate system. is the frame rate change parameter, The lower limit of the frame rate.

[0007] Furthermore, in step 2, the method for fusing the three-dimensional depth map after adjusting the frequency to obtain the local point cloud map is: Downsample the depth map according to the set sampling interval to extract the depth information of the pixel points in the depth map; calculate the spatial point position of each pixel point in the corresponding scene in the depth map in the camera coordinate system according to the camera intrinsic parameter matrix; Project the point cloud in the world coordinate system calculated in the current frame to the camera coordinate system of the previous frame, and then project the point cloud in the camera coordinate system of the previous frame to the camera pixel plane, and confirm that the spatial position of the common viewpoint in two adjacent frames remains unchanged, and the pixels at the corresponding positions in the previous frame of the camera have the same depth value; The UAV posture in the world coordinate system is estimated based on the visual odometer, and the transformation relationship between the multi-camera and the UAV posture in the front, left and right positions is obtained based on the camera calibration; based on the transformation relationship, the point cloud positions of the three cameras in the world coordinate system are calculated; Filter out point clouds whose errors are within the preset error threshold Construct a local point cloud map; i Indicates the camera number.

[0008] Furthermore, before clustering the point clouds corresponding to different objects in the local point cloud map in step 3, a collision probability attenuation strategy is also used to reduce the occupancy probability value of the grids that are not measured by the sensor within a defined time.

[0009] Furthermore, in step 3, a Euclidean distance-based clustering algorithm is used to cluster the point clouds corresponding to different objects in the transformed world coordinate system.

[0010] Furthermore, the method for constructing a grid map of a drone for fast detection of a target from multiple perspectives also includes updating the local area of ​​the grid using a ring buffer data structure, which is specifically described as follows: The array of the ring buffer data structure stores the occupancy probability value of each grid in the three-dimensional grid space and is offset grids, then the corresponding addressing relationship from the ring buffer to the occupied grid position is as described by the following formula: ; x is the grid coordinate, insideVolume(x) is used to determine whether the point represented by x is in the local map, address(x) is used to convert the grid coordinate into the index of the ring buffer data array, and mod is the remainder operation. N is the number of grids along each axis of the local map.

[0011] By occupying the position of the grid local map, the ring buffer position index corresponding to the occupied grid local map is calculated, and then the grid local map attribute is accessed to update the corresponding area. The update formula used is as follows: O k = O k-1 ( X k - X k-1 ) ; In the formula, X k =( i k ,i yk ,izk ) is the UAV in k The position of the grid map grid at the time, X k -X k-1 For drones k Time relative to k- The change in grid position at time 1, O k for k The offset of the time, O k-1 is the offset at time k-1.

[0012] Furthermore, the method for constructing a drone grid map for fast detection of follow-up targets from multiple perspectives also includes lightweight management of the drone grid map, which is specifically described as follows: The lightweight management of the drone grid map is reflected in the three parts: the area outside the local grid map area, the grid local map, and the new local area after the drone displacement change. For the area outside the local grid map area, the occupancy attribute is reset to 0, the occupancy probability attribute is reset to the initial value 0.12, and the access attribute is reset to 0; for the new local area after the drone displacement change, it is rebuilt according to steps 1-3; for any grid in the grid local map g l , when a point cloud appears in the previous frame but not in the current frame, the occupancy probability value is reduced; when the probability value is lower than the preset threshold, it is set to unoccupied.

[0013] The present invention provides a method for constructing a raster map of an unmanned aerial vehicle for rapid detection and following of targets from multiple perspectives. By fusing the environmental depth information perceived by the front, left, and right three groups of binocular cameras, a local point cloud map with three-way visual fusion is constructed. The space is rasterized, and the occupancy probability of each grid feature in each grid map is estimated, thereby constructing a local occupied grid map. Based on the map, a map dynamic information enhancement strategy is designed, and dynamic obstacle recognition and state estimation are realized based on information such as point cloud spatial distribution and map occupancy probability, thereby increasing the accuracy of dynamic obstacle construction in the local map. At the same time, in combination with the flight status of the unmanned aerial vehicle, a map construction strategy with adaptive frequency adjustment is designed to reduce unnecessary calculations while ensuring the accuracy of map construction. At the same time, in order to reduce the memory usage and calculation time used, a local map centered on the unmanned aerial vehicle is constructed based on the ring buffer data structure, and the map area range is maintained in real time.

[0014] In summary, the UAV grid map construction method of the present invention can provide safe and fast path planning for UAV target following tasks, and has the advantages of wide construction range, high accuracy, small amount of calculation, and small storage occupation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of the method for constructing a grid map of a drone in an embodiment; Figure 2 This is a schematic diagram of the necessary mapping area for the drone in the embodiment; Figure 3 A schematic diagram of dynamic object prediction in an embodiment; Figure 4 Schematic diagram of the grid local map update process in the embodiment; Figure 5 A 2D schematic diagram constructed for a local map. DETAILED DESCRIPTION

[0016] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is described in detail below in conjunction with the implementation modes and the accompanying drawings.

[0017] In this embodiment, the three binocular camera modules are installed in the front, left and right positions of the drone respectively, and the parameters are kept unchanged after setting the fixed focal length. The IMU, an important component of the flight control system, is installed in the center of the drone center board, and the ICM-20689 IMU module built into the PX4 flight control system is selected as a reference for binocular external parameter calibration.

[0018] like Figure 1-Figure 2 As shown, the present embodiment provides a method for constructing a drone grid map for rapid detection of a follow-up target from multiple perspectives, comprising the following steps: Step 1: Collect the flight status of the drone and the depth map of the surrounding environment in real time. The flight status includes the flight speed, airspeed, and yaw angle. The depth map of the surrounding environment is a three-way depth map captured by three binocular camera modules installed on the front, left, and right sides of the drone.

[0019] Step 2: Construct a static obstacle map based on the three-way depth map and flight status collected in step 1. The specific description is as follows: Multi-camera frequency adaptive mapping, this embodiment uses the flight status of the drone to adjust the frame rate of the cameras on the left and right sides of the drone in real time to adjust the mapping frequency in each direction and realize multi-camera frequency adaptive mapping. This operation not only ensures flight safety, but also reduces the amount of mapping calculation. The adjustment principle is described in the following formula: ; In the formula, is the frame rate of the left camera, is the frame rate of the camera on the right, The flight speed of the drone, is the flight speed direction of the drone; is the heading angle of the UAV, is the difference between the flight speed direction of the drone and the heading angle of the drone. The angles in this formula are defined in the spatial coordinate system; is the frame rate change parameter, The lower limit of the frame rate.

[0020] Binocular stereo matching, based on binocular matching to obtain the point cloud positions of three cameras in the world coordinate system is described as follows: The depth map is downsampled with a sampling interval of 2 pixels to extract the depth information of the pixel points in the depth map, as described in the following formula: ; K =1,2,3 0≤ u ≤640; In the formula, For the i The depth map of the camera is (u,v) The depth of the pixel location, Depth map (u,v) The depth value of the location in meters, I Indicates the number of point clouds in each camera coordinate system; f depth scaling is the depth scale factor, set to 1000.

[0021] According to the camera intrinsic parameter matrix, calculate the spatial point position in the camera coordinate system: ; In the formula, For the i The camera intrinsic parameter matrix; P ci Indicates i The coordinates of the point cloud in the camera coordinate system. n Indicates the number of array elements composed of the aforementioned point cloud, and does not represent the specific number of point clouds. j Indicates the first j A point cloud.

[0022] The pose of the drone IMU in the world coordinate system is estimated based on the visual odometer. The transformation relationship between the front, left and right multi-cameras and the drone IMU pose is obtained based on the camera calibration. The point cloud positions of the three cameras in the world coordinate system are calculated: ; In the formula, P wi is the spatial point cloud position of the three cameras in the world coordinate system; T wb It is the drone pose in the world coordinate system of the current frame. Tbci The transformation relationship between the i-th camera and the drone, P ci Indicates i The coordinates of the point cloud in the camera coordinate system, where i =1,2,3; when i =1, it indicates the transformation relationship between the forward multi-view camera and the drone obtained based on the camera calibration; when i =2, it indicates the transformation relationship between the left multi-camera and the drone obtained based on camera calibration; when i =3, it indicates the transformation relationship between the right multi-camera and the drone obtained based on camera calibration.

[0023] Point cloud filtering and fusion, projecting the point cloud in the world coordinate system calculated in the current frame to the camera coordinate system of the previous frame, and then projecting the point cloud in the camera coordinate system of the previous frame to the camera pixel plane. Filter out points whose errors are within the preset threshold , to construct a local point cloud map. The local point cloud map of this embodiment is as follows Figure 5 As shown. Project the point cloud in the world coordinate system calculated in the current frame to the camera coordinate system of the previous frame It is expressed as: ; In the formula, The drone pose in the world system of the previous frame, P wi The spatial point cloud position of which camera in the world coordinate system.

[0024] Then project the point cloud in the camera coordinate system of the previous frame onto the camera pixel plane and express it as: ; In the formula, Represents the mapping from two-dimensional image coordinates to three-dimensional camera coordinate system, is the reprojection depth value of the current point cloud, , is the reprojected image coordinate of the current point cloud, K ci is the intrinsic parameter matrix of the i-th camera, Represents the spatial position of the reprojected point cloud of the three cameras in the world coordinate system of the current frame.

[0025] Make the spatial position of the common viewpoint in two adjacent frames unchanged, and the corresponding position pixels in the previous frame of the camera have the same depth value, that is: ; In the formula, is the reprojection depth value of the current point cloud, The previous frame i The depth map of the camera is The depth of the pixel location, , is the reprojected image coordinate of the current point cloud. Considering factors such as sensor accuracy and the accuracy requirements of the map in the application scenario, this embodiment selects point clouds whose errors are within the preset error threshold. , forming a local point cloud map.

[0026] Build a local grid map, grid the space based on the local point cloud map and set the grid properties. The grid properties include occupancy properties , access properties and occupation probability. Occupancy attribute In the grid local map, there are three states: occupied 1, unoccupied 0, unknown -1, which are used to represent the obstacle information of the map; access properties It is used to avoid wasting computing resources by repeatedly visiting the same grid. It contains three states: occupied 1, visited -1, and not visited 0. hit It is used to describe the probability that a grid is an obstacle. The value range is 0.0~1.0. In order to increase the calculation efficiency, the occupancy probability is stored in logarithmic form, as described in the following formula: ; In the formula, For Grid The logarithmic occupancy probability, For Grid The probability of occupancy.

[0027] Considering the actual situation of drone map construction, in order to better construct the map, this embodiment expresses the posterior probability of the map as the product of the marginal probability under the condition that the drone posture and the observation point cloud are known, as described in the following formula: ; In the formula, g Represents the entire grid local map, Represents each grid, is the UAV pose sequence, is the depth camera observation sequence.

[0028] Initialize the local grid map in the range of 22m*22m*5m In the example, all grids are assigned initial data space and corresponding initial attributes, that is, the occupied attribute is 0, the access attribute is 0, and the occupancy probability is 0.12. , map the point cloud to grid locations: ; In the formula, is integer position data, is the grid local map resolution, in m / grid, by Each value of is rounded down, and the point cloud is finally divided into different grids.

[0029] In each frame, it may be an obstacle or a noise point. Based on the actual application characteristics of continuous multi-frame mapping, the occupancy probability of the point in each frame is increased to determine Whether it is an obstacle in multiple consecutive frames is described by the following formula: ; In the formula, For Grid The occupancy probability update value in the current frame, is the occupancy probability of the previous frame, is the probability increasing hyperparameter. In multiple consecutive frames, if If the probability exceeds the preset threshold, the point is considered The grid For obstacles, set the grid property to occupy 1.

[0030] In order to avoid the error parallax area from being enlarged into a block area, causing a large number of point clouds to gather in a grid and reduce the amount of calculation, this embodiment sets the access attribute to occupied 1 to ensure that each grid is only visited once in the single-frame mapping calculation. After the map is initialized, the occupancy probability value of the grid in the grid local map is calculated, and the grid local map is updated based on the probability value to obtain a static obstacle map.

[0031] Step 3: Construct a dynamic obstacle map based on the static obstacle map. The specific description is as follows: The collision probability attenuation strategy is adopted to reduce the influence of the previous occupancy probability value over time, and reduce the occupancy probability value of the grid that has not been measured by the sensor for a long time. The specific formula is as follows: ; In the formula, For Grid The coefficient of the occupancy probability decay over time. If there is no sensor measurement for a long time, the grid occupancy probability will converge to 0, that is, the idle state.

[0032] Local dynamic point cloud segmentation and target tracking state estimation, clustering objects in different states based on the occupancy probability value of the grid. Affected by the accuracy of binocular stereo matching, there are many high-frequency noise points in the depth estimation of the binocular camera at 5~8m, and such noise points will be estimated as dynamic points during probability estimation. Therefore, this embodiment is based on the Euclidean distance clustering method, first screening out grids with occupancy probabilities higher than the preset threshold, and then clustering the connectivity between each neighborhood grid according to certain connectivity rules within the neighborhood range with these grids as the center and a radius of 0.08 to obtain the position of the dynamic object. Each cluster position is represented as the average position of the grids contained in the cluster, that is, the cluster object. m The centroid , as described by the following formula:

[0033] ; In the formula, For clustering objects m The a points, n' is the clustering object m The total number of point sets in the cluster is based on the nearest neighbor matching method of the points to achieve clustering centroids between multiple frames At the same time, in order to ensure the robustness of dynamic estimation, the maximum moving speed of the dynamic object is considered to be 3m / s, and the depth estimation frequency is 8 frames. Under this condition, the nearest neighbor of the point between two frames should be less than 0.375m.

[0034] The speed of each cluster object is expressed as The moving speed in a multi-frame sequence is described by the following formula: ; In the formula, Is a clustering object m exist k The speed of time, For clustering objects m exist k The timestamp of the moment. From this process, the speed of each object clustered by the grid can be directly obtained. Since the speed of the object may change, the point cloud near the clustered object is resampled to achieve real-time adjustment of the object speed.

[0035] Following the target state prediction, based on the assumption of uniform speed of dynamic obstacles between two frames, the grid occupied by the dynamic point cloud is updated in the future time to compensate for the update delay of dynamic obstacles in the grid local map, such as Figure 3 shown.

[0036] Predict the obstacle position based on the speed and position prediction model: ; In the formula, dt for k Moment and k-1 The time difference between moments, express Normally distributed noise with zero mean, where is the covariance matrix. In the process of following the target speed estimation, the predicted position information of the obstacles is filtered according to the boundary conditions of the local grid map, and the dynamic point clouds whose predicted positions fall outside the map range are eliminated.

[0037] Step 4: Keep the drone positioning grid at the center grid. The center grid of this embodiment is determined by the sensor performance and the customized mapping range. Fix the direction of the map to the initial direction, update the grid local map according to the current flight status of the drone, and rebuild the grid local map according to steps 1 to 3 for the new local area after the displacement of the drone, until the target point is reached to complete the map construction and obstacle avoidance.

[0038] In the process of constructing the entire map, this embodiment uses a ring buffer data structure to store and update the grid local map, as well as lightweight management of the local grid map. The specific description of using the ring buffer data structure to store and update the grid local map is as follows: The array of ring buffer data structures stores the occupancy probability value of each grid in the grid local map and is offset grids, then the corresponding addressing relationship from the ring buffer to the occupied grid position is as described by the following formula: ; By occupying the position of the grid local map, the ring buffer position index corresponding to the occupied grid local map is calculated, and then the grid local map attribute is accessed to update the corresponding area. The update formula used is as follows: O k = O k-1 ( X k -X k-1 ) In the formula, X k =( i k ,i yk ,i zk ) is the UAV in k The position of the grid map grid at the time, X k -X k-1 For dronesk Time relative to k-1 The change in grid position at time, O k for k The offset of the time, O k-1 for k-1 The offset of the time.

[0039] As the drone flies in the direction of the arrow, the grids in the grid map area are removed in real time, and the grids in the newly generated grid map area are initialized. In order to complete this operation simply and quickly, the ring buffer index of the removed grid is assigned to the grid that newly enters the local map and initialized. In this way, the grid map can be updated at high speed according to the movement of the drone.

[0040] like Figure 4 As shown in the figure, the specific description of the lightweight management of the local grid map is: The lightweight management of the drone grid map is reflected in the three parts: the area outside the local grid map area, the grid local map, and the new local area after the drone displacement change. For the area outside the local grid map area, the occupancy attribute is reset to 0, the occupancy probability attribute is reset to the initial value 0.12, and the access attribute is reset to 0; for the new local area after the drone displacement change, it is rebuilt according to steps 1-3; for any grid in the grid local map , when a point cloud appears in the previous frame but not in the current frame, the occupancy probability is reduced.

[0041] In addition, in order to increase the search space for subsequent planning, the ray casting algorithm is used to project the ray path to the drone from any point in the space after reprojection. The grid on the projection path will reduce the occupancy probability value. visit , set the grid that passes through the ray casting path to -1, and then skip other points when casting light to the grid visit The grid with attribute -1 avoids repeated calculation of grids close to the drone.

Claims

1. A method for constructing a raster map of an unmanned aerial vehicle for rapid detection and following of targets from multiple perspectives, characterized in that: The following steps are involved: Step 1: Collect the flight status of the drone and the depth map of the surrounding environment in real time. The flight status includes the flight speed, airspeed, and yaw angle. The depth map of the surrounding environment is a three-way depth map perceived by binocular cameras installed in the front, left, and right directions of the drone. Step 2: dynamically adjust the frequency of the three-dimensional depth map information according to the flight status of the UAV, fuse the three-dimensional depth map after the frequency adjustment to obtain a local point cloud map; rasterize the space based on the local point cloud map to obtain a grid local map; use the grid posterior probability algorithm to calculate the probability value of each grid in the grid local map, and use the probability value to update the occupied grid local map to obtain a static obstacle map; Step 3: Cluster the point clouds corresponding to different objects in the local point cloud map to obtain the position of the dynamic object; estimate the actual position of the dynamic object based on the uniform speed assumption of the dynamic obstacle between the two frames of the depth map; predict the grid position of the corresponding dynamic point cloud mapped to the static obstacle map according to the actual position of the dynamic object, and calculate its probability value to update the static obstacle map to obtain the dynamic obstacle map; Step 4: Keep the drone positioning grid at the center grid, fix the direction of the map to the initial direction, update the occupied grid local map according to the current flight status of the drone, and rebuild the grid local map according to steps 1-3 for the new local area after the drone displacement change until the target point is reached to complete navigation and obstacle avoidance.

2. The method for constructing a raster map of an unmanned aerial vehicle for rapid detection of a follow-up target from multiple perspectives according to claim 1, characterized in that: The principle of dynamically adjusting the frequency of the three-way depth map information according to the flight status of the drone in step 2 is: ； In the formula, is the frame rate of the left camera, is the frame rate of the camera on the right, The flight speed of the drone, is the flight speed direction of the drone; is the heading angle of the UAV, is the difference between the UAV flight speed direction and the UAV heading angle. The angles in this formula are defined in the spatial coordinate system. is the frame rate change parameter, The lower limit of the frame rate.

3. The method for constructing a UAV grid map for rapid detection of follow-up targets from multiple perspectives according to claim 2, characterized in that: The method for fusing the three-dimensional depth map after adjusting the frequency in step 2 to obtain a local point cloud map is: Downsample the depth map according to the set sampling interval to extract the depth information of the pixel points in the depth map; calculate the spatial point position of each pixel point in the corresponding scene in the depth map in the camera coordinate system according to the camera intrinsic parameter matrix; The UAV posture in the world coordinate system is estimated based on the visual odometer, and the transformation relationship between the multi-camera and the UAV posture in the front, left and right positions is obtained based on the camera calibration; based on the transformation relationship, the point cloud positions of the three cameras in the world coordinate system are calculated; Project the point cloud in the world coordinate system calculated in the current frame to the camera coordinate system of the previous frame, and then project the point cloud in the camera coordinate system of the previous frame to the camera pixel plane; Confirm that the spatial position of the common viewpoint in two adjacent frames remains unchanged, and the pixels at the corresponding position in the previous frame of the camera have the same depth value; Filter out point clouds whose errors are within the preset error threshold , forming a local point cloud map; i Indicates the camera number.

4. The method for constructing a UAV grid map for rapid detection of follow-up targets from multiple perspectives according to claim 3 is characterized in that: Before clustering the point clouds corresponding to different objects in the local point cloud map in step 3, a collision probability attenuation strategy is also used to reduce the occupancy probability value of the grids that are not measured by sensors within a defined time.

5. The method for constructing a drone grid map for rapid detection of follow-up targets from multiple perspectives according to claim 4, characterized in that: In step 3, a Euclidean distance-based clustering algorithm is used to cluster the point clouds corresponding to different objects in the transformed world coordinate system.

6. The method for constructing a UAV grid map for multi-view rapid detection of follow-up targets according to claim 1, characterized in that: The method for constructing a grid map of a drone for fast detection of a follow-up target from multiple perspectives also includes updating a local grid map using a ring buffer data structure, which is specifically described as follows: The array of ring buffer data structures stores the occupancy probability value of each grid in the grid local map and is offset grids, the offset defines the subsequent corresponding addressing relationship from the ring buffer to the occupied grid position, as described by the following formula: ； Among them, x is the grid coordinate, insideVolume(x) is to determine whether the point represented by x is in the local map, address(x) is the index of the ring buffer data array converted from the grid coordinate, mod is the remainder operation, N is the number of grids along each axis of the local map; By occupying the position of the grid local map, the ring buffer position index corresponding to the occupied grid local map is calculated, and then the grid local map attribute is accessed to update the corresponding area. The update formula used is as follows: O k = O k-1 ( X k -X k-1 ) ； In the formula, X k =( i k ,i yk ,i zk ) is the UAV in k The position of the grid on the raster map where the moment is located, X k -X k-1 For drones k Time relative to k-1 The change in the grid position at time is k The offset of the time, O k-1 for k-1 The offset of the time.

7. The method for constructing a UAV grid map for multi-view rapid detection of follow-up targets according to claim 1, characterized in that: The method for constructing a drone grid map for fast detection of follow-up targets from multiple perspectives also includes lightweight management of the drone grid map, which is specifically described as follows: The lightweight management of the drone grid map is reflected in three parts: the area outside the local grid map area, the grid local map, and the new local area after the drone displacement changes; for the area outside the local grid map area, the occupancy attribute is reset to 0, the occupancy probability attribute is reset to the initial value of 0.12, and the access attribute is reset to 0; for any grid in the grid local map g l ,When a point cloud appears in the previous frame but not in the current frame, the occupancy probability value is reduced.

Images