Point cloud sequence data processing method and device, computer device and storage medium

Abstract

The application relates to a point cloud sequence data processing method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring point cloud sequence data of an area where a moving object is located in a target scene, locating obstacle points of each frame of point cloud sequence in the point cloud sequence data, projecting the obstacle points of each frame of point cloud sequence to obtain occupancy grid map sequence data, each occupancy grid map carrying a time parameter and position information of the moving object when the occupancy grid map is constructed, correcting the occupancy grid map at the current time based on the time parameter and the position information of the moving object, and determining the motion obstacle points of the occupancy grid map at the current time and the motion information of the motion obstacle points. According to the scheme, multi-frame point cloud data is processed, and an accurate occupancy grid map at the current time can be obtained, so that the accuracy of the drivable area detection is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer device, storage medium, and computer program product for processing point cloud sequence data. Background Technology

[0002] Autonomous navigation of unmanned vehicles or intelligent robots requires accurate detection of obstacles in the robot's area and the provision of a grid map showing obstacle occupancy. Therefore, this presents a challenge to the detection requirements of drivable road surface detection solutions.

[0003] In traditional road drivable area detection technology, single-frame LiDAR is generally used for detection. Appropriate algorithms are used to locate ground points in the scene point cloud. Then, two-dimensional projection is performed on the obstacle points to give an occupancy grid map of the area. For the extraction of ground points, traditional algorithms use manually made feature representations of the ground, which have poor generalization and perform poorly in complex scenes such as the wild, bridges, and tunnels, and cannot truly reflect scene information.

[0004] Current deep learning-based methods generally use single-frame point clouds as input. Due to the sparsity of the point cloud itself and the occlusion of obstacles, even if the ground point cloud and obstacle points are extracted, the grid map cannot fully reflect the real information of the scene, resulting in low detection accuracy of the drivable area of ​​the road.

[0005] In summary, existing detection methods for drivable road surfaces suffer from low detection accuracy. Summary of the Invention

[0006] Therefore, it is necessary to provide a point cloud sequence data processing method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the detection accuracy of drivable areas of road surfaces, addressing the aforementioned technical problems.

[0007] Firstly, this application provides a method for processing point cloud sequence data. The method includes:

[0008] Acquire point cloud sequence data of the area where the moving object is located in the target scene;

[0009] Locate obstacle points in each frame of the point cloud sequence data;

[0010] The obstacle points of each frame point cloud sequence are projected to obtain the occupied grid map sequence data. Each occupied grid map carries time parameters and the position information of the moving object when the occupied grid map is constructed.

[0011] Based on time parameters and the position information of moving objects, the current occupied grid map is corrected, and the moving obstacle points in the current occupied grid map and their movement information are determined.

[0012] In one embodiment, the position information of the moving object includes the position information of the moving object in the global coordinate system;

[0013] Based on time parameters and the position information of moving objects, the correction of the current occupied grid map includes:

[0014] Based on the position information of moving objects in the global coordinate system, the occupied raster map sequence data is transformed to the global coordinate system;

[0015] Extract multiple target occupancy raster maps from the occupancy raster map sequence data after coordinate transformation. The target occupancy raster map is the previous occupancy raster map that overlaps with the occupancy raster map at the current moment in the area where the moving object is located.

[0016] Obstacle clustering is performed on multiple target-occupied grid maps to obtain obstacle blocks for each target-occupied grid map;

[0017] Obtain the intersection-over-union ratio (IoU) of obstacle blocks in the occupied grid map at the current time. The IoU is the overlap rate between obstacle blocks in the occupied grid map at the current time and obstacle blocks in the target occupied grid map.

[0018] Based on the intersection-union ratio of obstacle blocks in the current occupied grid map, the current occupied grid map is corrected, and the moving obstacle points and their motion information are determined in the current occupied grid map.

[0019] In one embodiment, correcting the occupancy grid map at the current moment based on the intersection-union ratio of each obstacle block in the occupancy grid map at the current moment includes:

[0020] If the intersection-union ratio of each obstacle block in the current occupied grid map is greater than or equal to the first preset ratio threshold, then the missing obstacles in the current occupied grid map are supplemented to correct the current occupied grid map. The missing obstacles are those recorded in the target occupied grid map but not in the current occupied grid map.

[0021] In one embodiment, based on the intersection-over-union ratio of each obstacle block in the occupancy grid map at the current moment, determining the moving obstacle points in the occupancy grid map at the current moment and the motion information of the moving obstacle points includes:

[0022] If the intersection-union ratio of the obstacle block occupying the grid map at the current moment is less than the second preset ratio threshold, then the obstacle block is identified as a moving obstacle point, and the second preset ratio threshold is less than the first preset ratio threshold.

[0023] Obtain the first center coordinates of the obstacle point in the current time-occupied grid map, and the second center coordinates of the target in the previous time-occupied grid map;

[0024] The displacement of the obstacle point is obtained based on the first center coordinate and the second center coordinate;

[0025] The speed of the obstacle is determined based on its displacement and the time difference between the current moment and the previous moment.

[0026] In one embodiment, the obstacle points in the location point cloud sequence data include:

[0027] Feature encoding is performed on the point cloud sequence data to obtain point feature vectors;

[0028] Extract local features from the point feature vectors and map these local features onto the original scene to obtain scene feature data;

[0029] Decode the scene feature data to obtain the scene elevation feature value;

[0030] Based on the scene elevation feature value and the coordinate value of the point cloud sequence data, locate the obstacle point in the point cloud sequence data.

[0031] In one embodiment, feature encoding of point cloud sequence data to obtain point feature vectors includes:

[0032] Each frame of point cloud sequence data in the point cloud sequence data is divided into several cylindrical regions;

[0033] Select target cylindrical regions from several cylindrical regions to find the number of point clouds that meets the preset requirements;

[0034] The RANSAC (Random Sample Consensus) algorithm is used to cluster the point cloud data in each target cylindrical region into a plane. The point cloud data of the target cylindrical region is then augmented with features based on the normal of the plane to obtain the point feature vector of the target cylindrical region.

[0035] Secondly, this application also provides a point cloud sequence data processing apparatus. The apparatus includes:

[0036] The data acquisition module is used to acquire point cloud sequence data of the area where the moving object is located in the target scene. The specified area is a preset distance range centered on the current position information of the moving object.

[0037] The obstacle point localization module is used to locate obstacle points in each frame of the point cloud sequence data.

[0038] The occupancy grid map data acquisition module is used to project obstacle points into each frame point cloud sequence to obtain occupancy grid map sequence data. Each occupancy grid map carries time parameters and the position information of moving objects when constructing the occupancy grid map.

[0039] The Occupied Grid Map Correction Module is used to correct the current occupied grid map based on time parameters and the position information of moving objects, and to determine the moving obstacle points and their movement information in the current occupied grid map.

[0040] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0041] Acquire point cloud sequence data of the area where the moving object is located in the target scene;

[0042] Locate obstacle points in each frame of the point cloud sequence data;

[0043] The obstacle points of each frame point cloud sequence are projected to obtain the occupied grid map sequence data. Each occupied grid map carries time parameters and the position information of the moving object when the occupied grid map is constructed.

[0044] Based on time parameters and the position information of moving objects, the current occupied grid map is corrected, and the moving obstacle points in the current occupied grid map and their movement information are determined.

[0045] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0046] Acquire point cloud sequence data of the area where the moving object is located in the target scene;

[0047] Locate obstacle points in each frame of the point cloud sequence data;

[0048] The obstacle points of each frame point cloud sequence are projected to obtain the occupied grid map sequence data. Each occupied grid map carries time parameters and the position information of the moving object when the occupied grid map is constructed.

[0049] Based on time parameters and the position information of moving objects, the current occupied grid map is corrected, and the moving obstacle points in the current occupied grid map and their movement information are determined.

[0050] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0051] Acquire point cloud sequence data of the area where the moving object is located in the target scene;

[0052] Locate obstacle points in each frame of the point cloud sequence data;

[0053] The obstacle points of each frame point cloud sequence are projected to obtain the occupied grid map sequence data. Each occupied grid map carries time parameters and the position information of the moving object when the occupied grid map is constructed.

[0054] Based on time parameters and the position information of moving objects, the current occupied grid map is corrected, and the moving obstacle points in the current occupied grid map and their movement information are determined.

[0055] The aforementioned point cloud sequence data processing method, apparatus, computer equipment, storage medium, and computer program product differ from traditional technologies that use single-frame point cloud data to represent scene information. Instead, it uses multi-frame point cloud data, i.e., point cloud sequence data, as input. It locates obstacle points in each frame of the point cloud sequence data and projects these obstacle points to obtain a multi-frame point cloud data occupancy grid map sequence. Then, based on the time parameters and moving object position information in the occupancy grid map, it corrects the current occupancy grid map and determines the moving obstacle points and their motion information. This entire scheme uses multi-frame point cloud data for processing to generate a more accurate occupancy grid map. Furthermore, correcting the current occupancy grid map further improves its accuracy. Additionally, determining the moving obstacle points and their motion information significantly improves the accuracy of drivable area detection. Attached Figure Description

[0056] Figure 1 This is an application environment diagram of a point cloud sequence data processing method in one embodiment;

[0057] Figure 2 This is a flowchart illustrating a point cloud sequence data processing method in one embodiment;

[0058] Figure 3 This is a flowchart illustrating the steps of correcting the occupied grid map and determining the moving obstacle points and their motion information in one embodiment.

[0059] Figure 4This is a detailed flowchart illustrating the steps of correcting the occupied grid map and determining the moving obstacle points and their motion information in another embodiment;

[0060] Figure 5 This is a detailed flowchart illustrating the process of locating obstacle points in a point cloud sequence data in one embodiment.

[0061] Figure 6 This is a detailed flowchart illustrating a point cloud sequence data processing method in one embodiment;

[0062] Figure 7 This is a structural block diagram of a point cloud sequence data processing device in one embodiment;

[0063] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0065] The point cloud sequence data processing method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, the moving object 102 connects to the server 104 and the lidar 106 (which can be mounted on the moving object 102) via a network. Figure 1(Not specified in the text) communicates. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104 or placed on the cloud or other network servers. Specifically, taking a robot as an example, and a road as an example, the moving object 102 can be a robot driving on a road. The LiDAR 106 collects point cloud sequence data of the area where the robot 102 is located and sends the collected point cloud sequence data to the server 106. After receiving the point cloud sequence data, the server 106 locates the obstacle points in each frame of the point cloud sequence data. Then, it projects the obstacle points in each frame of the point cloud sequence to obtain the occupancy grid map sequence data. Each occupancy grid map carries time parameters and the position information of the moving object when the occupancy grid map is constructed. Based on the time parameters and the position information of the moving object, the occupancy grid map at the current moment is corrected, and the moving obstacle points and the movement information of the moving obstacle points in the occupancy grid map at the current moment are determined. The occupancy grid map carrying the movement information of the moving obstacle points is sent to the navigation module of the robot 102 so that the robot 102 can plan its driving path. The moving object 102 can be, but is not limited to, various mobile robots, vehicles, and drones. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0066] In one embodiment, such as Figure 2 As shown, a point cloud sequence data processing method is provided, which can be applied to... Figure 1 Taking server 104 as an example, the following steps are included:

[0067] Step 100: Obtain point cloud sequence data of the area where the moving object is located in the target scene.

[0068] The target scene can be urban roads, outdoor roads, mountainous areas, or fields. The target object can be a moving robot, a vehicle, or other moving objects. Point cloud sequence data refers to a sequence of multiple frames of point cloud data that are consecutive in time; that is, a point cloud sequence includes multiple frames of point cloud data within a certain time range. In this embodiment, the target scene is illustrated using an urban road as an example, and the moving object is illustrated using a robot as an example.

[0069] In practical applications, the first step is to perform information calibration to obtain the robot's position and orientation information in the global coordinate system. Specifically, this involves calibrating the LiDAR, GPS (Global Positioning System), and IMU (Inertial Measurement Unit) to ensure that they can synchronously provide corresponding sensor data and timestamp information. Next, global and local coordinate systems for the robot's motion are established. The global coordinate system is denoted as H, with the initial position as the origin, creating a northeast-sky global coordinate system. East is the x-axis, north is the y-axis, and the upward direction is the z-axis. The local coordinate system is the vehicle-mounted LiDAR coordinate system, denoted as L, with the area directly in front of the LiDAR as the x-axis, the area directly to the left as the y-axis, and the upward direction as the z-axis. The LiDAR's position is adjusted so that its coordinate system completely coincides with the robot's coordinate system. The LiDAR will provide point cloud data P' in the local coordinate system. By fusing the GPS and IMU sensor data, the robot's position and orientation information in the global coordinate system can be obtained. Then, the lidar collects point cloud sequence data of the area where the robot is located, based on the robot's position and orientation information.

[0070] Step 200: Locate the obstacle points in each frame of the point cloud sequence data.

[0071] Obstacle points and ground points are relative terms; that is, the point cloud data remaining after removing ground points can be considered as obstacle points. In this embodiment, after acquiring the point cloud sequence data of the robot's location area, before locating the obstacle points in each frame of the point cloud sequence data, point cloud preprocessing is required. Specifically, point cloud preprocessing includes coordinate correction, point cloud filtering, and the removal of invalid point cloud data to obtain valid point cloud data. Specifically, for the point cloud data P' provided by the LiDAR, the robot's orientation information provided by the positioning system can be used to correct the coordinates of the point cloud. The correction process for point cloud P' can be as follows: Let the coordinates of any point in P' be (x', y', z'), and the position information Gi under the same timestamp of P' be represented by (xi, yi, zi, ai, bi, ci, di). Rotate the point cloud, and the robot's orientation quaternion is q = a0 + b0*i + c0*j + d0*k. Let p = x'*i + y'*j + z'*k. The rotated pure quaternion R(p) = q^-1*p*q, with its imaginary part representing the rotated point cloud coordinates (x, y, z), is then used. A point cloud filter is applied to extract points with x-coordinates at (-Δx, Δx) and y-coordinates at (-Δy, Δy), while invalid points are removed. This ensures that valid scene point clouds at corresponding locations can be extracted for each frame of the point cloud sequence. Finally, ground points are removed from each frame of the point cloud sequence data after preprocessing to obtain the obstacle points for each frame.

[0072] Step 300: Project the obstacle points of each frame point cloud sequence to obtain the occupied grid map sequence data. Each occupied grid map carries time parameters and the position information of the moving object when the occupied grid map is constructed.

[0073] The occupancy raster map sequence data consists of the occupancy raster map corresponding to each frame of the point cloud sequence. The occupancy raster map is also called the occupied raster map. It is a map description method where the space is divided into a finite number of graticles mi, the grid side length representing the division precision, and each grid is described by its occupancy probability P and coordinates. Continuing with the above embodiment, after locating the obstacle points in each frame of the point cloud sequence, a two-dimensional projection can be performed on the obstacle points in each frame to obtain the occupancy raster map corresponding to each frame of the point cloud. Specifically, this can involve obtaining the region (where the robot is located) with x-coordinates in (-Δx, Δx) and y-coordinates in (-Δy, Δy). At a given resolution r, the point cloud data of this region is rasterized to obtain a two-dimensional raster map. This raster map is represented by a two-dimensional array. The number of obstacle points falling into each grid cell is calculated. If the number of point cloud points falling into a grid cell is greater than a threshold T, then that grid cell is considered an obstacle grid cell, and the array value can be recorded as 100 to indicate the presence of an obstacle in the region; otherwise, the array value for that region is recorded as 0 to indicate that the region is passable. L raster map information is saved. If there are more than L pieces of information, the raster map information with the earliest timestamp is deleted. L can be set to different sizes according to the robot's travel speed. The occupied raster map at each moment can be obtained through the above method, resulting in an occupied raster map sequence, where each raster map carries both timestamp information and position parameters. Then, the occupied grid map can be represented by a sequence (x, y, t, M), where x and y represent the robot's position when the grid map is obtained, t is the time corresponding to the grid map, and M is the grid number. Figure 2 A dimensional array can be used to represent the distribution of obstacles around a robot.

[0074] Step 400: Based on the time parameters and the position information of the moving object, correct the current occupied grid map, and determine the moving obstacle points in the current occupied grid map and the movement information of the moving obstacle points.

[0075] A moving obstacle point is an obstacle point that has been displaced and whose movement speed is significantly greater than the movement speed of the grid map. Motion information includes the speed and direction of movement. In this embodiment, after obtaining multiple spatiotemporal occupied grid maps, the occupied grid map at the current time can be corrected using the occupied grid map within a time interval Δt before time t, and the motion information of the obstacles can be further supplemented. Specifically, based on the time parameters of the occupied grid map and the robot's position information, the occupied grid map at the current time can be corrected, and the moving obstacles and their motion information can be determined. The determination of moving obstacle points can be achieved by listing the grids where obstacles are located in each frame in spatiotemporal order, comparing the occupied grid maps of adjacent frames, and separating the moving obstacles. The direction and speed of the moving obstacles are determined based on the point cloud sequence stream. If the movement speed of the same obstacle is significantly greater than the movement speed of the grid map, it is considered a moving obstacle point, and its speed and direction are marked. Furthermore, the corrected occupied grid map at the current time, marked with the motion information of the moving obstacles, can be published and provided to the rare path planning and navigation modules.

[0076] The above-described point cloud sequence data processing method differs from traditional techniques that use single-frame point cloud data to represent scene information. Instead, it uses multi-frame point cloud data (i.e., point cloud sequence data) as input, locating obstacle points in each frame of the point cloud sequence. By projecting these obstacle points, it obtains a multi-frame point cloud data occupancy grid map sequence. Then, based on the time parameters and moving object position information in the occupancy grid map, it corrects the current occupancy grid map and determines the moving obstacle points and their motion information. This entire scheme, using multi-frame point cloud data to generate a multi-frame point cloud data occupancy grid map, results in a more accurate occupancy grid map. Furthermore, correcting the current occupancy grid map further improves its accuracy. Additionally, determining the moving obstacle points and their motion information significantly enhances the accuracy of drivable area detection.

[0077] like Figure 3 As shown, in one embodiment, the position information of the moving object includes the position information of the moving object in the global coordinate system;

[0078] Based on time parameters and the position information of moving objects, the correction of the current occupied grid map includes:

[0079] Step 402: Based on the position information of the moving object in the global coordinate system, the occupied grid map sequence data is converted to the global coordinate system.

[0080] Step 404: Extract multiple target occupancy grid maps from the occupancy grid map sequence data after coordinate transformation. The target occupancy grid map is the previous occupancy grid map that overlaps with the occupancy grid map at the current moment in the area where the moving object is located.

[0081] Step 406: Perform obstacle clustering on multiple target-occupied grid maps to obtain obstacle blocks for each target-occupied grid map.

[0082] Step 408: Obtain the intersection-union ratio (IUU) of obstacle blocks in the occupied grid map at the current time. The IUU is the overlap rate between the obstacle blocks in the occupied grid map at the current time and the obstacle blocks in the target occupied grid map.

[0083] Step 410: Based on the intersection-union ratio of obstacle blocks in the current occupied grid map, correct the current occupied grid map, and determine the moving obstacle points and their motion information in the current occupied grid map.

[0084] The intersection-union ratio (IUGR) is the ratio of the intersection to the union of two bounding boxes. In this embodiment, the IUGR of obstacle blocks in the current occupied grid map refers to the ratio of the number of grids occupied by each obstacle block in the current occupied grid map to the number of grids occupied by a certain target occupied grid map. The robot's position information is its position information in the global coordinate system. The process of correcting the occupied grid map at the current moment can be as follows: Since each occupied grid map contains the robot's position information (x, y) in the global coordinate system when the occupied grid map was obtained, for each occupied grid map in the occupied grid map sequence, this coordinate (x, y) can be used to perform coordinate system alignment of these L grid maps in the global coordinate system, that is, to transform the occupied grid map sequence data to the global coordinate system. Then, let the current time be tn, the current robot position be (xn, yn), and the occupancy grid map at the current time be Mn. Mn reflects the obstacle distribution in the region where the x-coordinate is in (xn-Δx, xn+Δx) and the y-coordinate is in (yn-Δy, yn+Δy). If the robot is at time ti before tn, and at that time the robot's xi coordinate is in (xn-Δx, xn+Δx) and its y coordinate is in (yn-Δy, yn+Δy), then the occupancy grid map Mi obtained at time ti also contains some information of Mn. Therefore, Mi is determined as the target occupancy grid map, and Mn is corrected using Mi. In this embodiment, it can be: if the coordinates (xi, yi) of the occupancy grid map in the global coordinate system at a certain time satisfy the following relationship, then the occupancy grid map is determined as the target occupancy grid map:

[0085] xn-2*Δx <xi<xn+2*Δx

[0086] yn-2*Δy <yi<yn+2*Δy

[0087] Extract all the occupied grid maps that satisfy the above inequality group as the target occupied grid maps, which are used to subsequently correct the occupied grid map at the current moment. Finally, p target occupied grid maps are extracted as the target occupied grid map set Mp (p < L). Then, for each target occupied grid map, obstacle clustering is performed. That is, for each obstacle grid, if the adjacent area also contains obstacle grids, then these two obstacle grids are regarded as the same obstacle and clustered into an obstacle block. Obstacle clustering is performed on each target occupied grid map in the above manner to obtain the obstacle blocks of each target occupied grid map. Then, taking the occupied grid map at the current moment as Mn and the previous target occupied grid map set as Mp, pairwise comparison is performed between Mn and each target occupied grid map in Mp, and Mn is corrected. The comparison process is as follows:

[0088] According to the result of the previous coordinate system alignment, Mn and Mp are mapped to the global coordinate system. For each obstacle block O in Mn, calculate the intersection over union of the occupied grid numbers of this obstacle clustering in the Mn map and the Mp map. According to the intersection over union, correct the occupied grid map at the current moment and determine the moving obstacle points and the motion information of the moving obstacle points in the occupied grid map at the current moment.

[0089] As Figure 4 shown, based on the intersection over union of the obstacle blocks in the occupied grid map at the current moment, correcting the occupied grid map at the current moment includes: Step 411, if the intersection over union of each obstacle block in the occupied grid map at the current moment is greater than or equal to the first preset ratio threshold, then supplement the missing obstacles in the occupied grid map at the current moment to correct the occupied grid map at the current moment. The missing obstacles are the obstacles recorded in the target occupied grid map but not recorded in the occupied grid map at the current moment.

[0090] If the intersection-union ratio (IU) is greater than or equal to the set first proportional threshold γ, then obstacle block O can be considered a stationary obstacle point, and the information captured by Mp and Mn for this obstacle block is very accurate. Mp's information can be used to correct Mn, that is, to locate obstacles identified in the target occupancy grid map at a previous time step but not identified in the current time step, and record them as missing obstacles. Specifically, if Mp records a region as an obstacle point, but Mn does not record it, then Mn supplements the missing obstacle information, records the region as an obstacle point, and assigns the array value of this region an integer greater than 0 but less than 100, indicating that an obstacle point may exist. This value is determined by the value of tn-ti; if the time interval is large, this value is smaller. If the IU is less than the first proportional threshold γ, it indicates that this obstacle is either a moving obstacle point, or the information recorded by Mp or Mn about the obstacle point is not accurate enough; for example, the obstacle point exists in part Mi but is completely present in Mn. In this case, Mp is not used to correct Mn. After correcting the current occupied grid map, the moving obstacle points and their motion information can be further determined based on the intersection-union ratio (IU). In this embodiment, through a series of processes such as coordinate system alignment, obstacle clustering, and IU calculation, the current occupied grid map can be corrected based on the previous occupied grid map to improve and supplement its information, and to accurately locate the moving obstacle points.

[0091] like Figure 4 As shown, in one embodiment, determining the moving obstacle points and their motion information in the current occupied grid map based on the intersection-over-union ratio of each obstacle block in the current occupied grid map includes:

[0092] Step 412: If the intersection-union ratio of the obstacle block occupying the grid map at the current moment is less than the second preset ratio threshold, then the obstacle block is identified as a moving obstacle point, and the second preset ratio threshold is less than the first preset ratio threshold.

[0093] Step 414: Obtain the first center coordinates of the obstacle point in the current time-occupied grid map, and the second center coordinates of the target in the previous time-occupied grid map.

[0094] Step 416: Obtain the displacement of the moving obstacle point based on the first center coordinates and the second center coordinates.

[0095] Step 418: Determine the speed of the obstacle based on its displacement and the time difference between the current moment and the previous moment.

[0096] Following the previous embodiment, if the intersection-union ratio (IUR) of an obstacle block in the currently occupied grid map is less than a first proportional threshold γ and less than a second preset proportional threshold β, then the obstacle block is identified as a moving obstacle point, and the moving obstacle point is labeled in the currently occupied grid map based on the previous frame's occupied grid map. Specifically, let Mn-1 be the previous frame sequence of Mn, and we label Mn as a moving obstacle point based on Mn-1. Since the algorithm is highly efficient, theoretically, Mn-1 and Mn represent obstacle points in essentially the same way. If the ratio of the number of overlapping grid cells of the obstacle block in the currently occupied grid map to the total number of obstacle cells (i.e., the IUR) is extremely low, less than the first proportional threshold γ and less than the second preset proportional threshold β, then this obstacle is considered a moving obstacle point. Then, calculate the first center coordinates (the average of the x and y coordinates of the obstacle point) of the moving obstacle point in Mn-1 and the second center coordinates under Mn. Then, calculate the displacement of the moving obstacle point based on the difference between the first and second center coordinates, and calculate the moving speed of the moving obstacle point based on the time difference between Mn-1 and Mn, and mark it under the Mn grid.

[0097] In this embodiment, the moving obstacle point is determined based on the relationship between the intersection-union ratio of the obstacle blocks in the current occupied grid map and the second preset ratio threshold. The moving speed of the moving obstacle point is determined based on the center coordinates of the moving obstacle point in the current occupied grid map and the occupied grid map in the previous time, which is both simple and accurate.

[0098] like Figure 5 As shown, in one embodiment, step 200 includes:

[0099] Step 220: Perform feature encoding on the point cloud sequence data to obtain point feature vectors.

[0100] Step 240: Extract local features from the point feature vectors and map the local features onto the original scene to obtain scene feature data.

[0101] Step 260: Decode the scene feature data to obtain the scene elevation feature value.

[0102] Step 280: Based on the scene elevation feature value and the coordinate value of the point cloud sequence data, locate the obstacle point in each frame of the point cloud sequence data.

[0103] In this embodiment, locating obstacle points in the point cloud sequence data can be achieved by: performing feature encoding on each frame of the point cloud sequence data, and performing feature augmentation on the point features to obtain a high-dimensional point feature vector, taking D-dimensional features as an example. Then, the PointNet feature extraction module is applied to the D-dimensional features of each point to extract point cloud features, and then a max pooling operation is applied to the point cloud features to obtain their local features, resulting in the local features of the point feature vector. The local features are then mapped to the original scene according to the coordinates at the time of extraction to obtain scene feature data. This scene feature data can be a pseudo-image of size (C, H, W), where C represents the local features, and H and W represent the length and width of the scene. This pseudo-image can represent the features of the entire scene.

[0104] Next, the pseudo-image is used as input to the SegNet network to regress and obtain the elevation feature values ​​of the scene. These feature values ​​reflect the ground elevation features of the region where each point cloud is located. After decoding, for each point in the point cloud, the network provides its elevation feature value and compares it with the z-value of the point. If the z-value is greater than the elevation feature value, the point is identified as an obstacle; if the z-value is less than the elevation feature value, the point is identified as a ground point. Following this method, all ground points in the scene can be separated, and the remaining points are considered obstacles. The SegNet network is an encoder-decoder network. The encoder consists of four convolutional layers. After every two convolutional layers, max pooling is performed using a 2×2 window and a stride of 2 (non-overlapping window). The encoder output is second-sampled using a factor of 4 (i.e., a 4×4 window and a stride of 4). Each encoder layer has a corresponding decoder layer, therefore the decoder network has four convolutional layers, with max unpooling performed after every two convolutional layers. The final decoder output has the same shape as the input pseudo-image (grid) and is fed into a 3x3 convolutional filter. This filter regresses the ground elevation value of each pixel in the pseudo-image (cells in the grid) and provides a feature value. The loss function of the neural network is defined as... Where I represents the true ground elevation map of a grid with shape (H, W), and ^I represents the predicted elevation map. Hyperparameters α and β are used to balance these two losses: and Specifically, the gradients in the x and y directions of the elevation view. It should be noted that in this embodiment, the neural network can be trained using the semantic Kitti dataset until convergence.

[0105] In this embodiment, by encoding and then decoding the point cloud data, the elevation feature value of each point cloud is obtained. By comparing the elevation feature value and z-value of the point cloud, the ground points in the scene can be separated simply and efficiently, and obstacle points can be screened out.

[0106] In one embodiment, step 220 includes: dividing each frame of point cloud sequence data in the point cloud sequence data into several cylindrical regions; selecting target cylindrical regions from the several cylindrical regions whose point cloud quantity meets a preset requirement; using the RANSAC algorithm to cluster and fit the point cloud data in each target cylindrical region into a plane; and performing feature augmentation on the point cloud data of the target cylindrical region according to the normal of the plane to obtain the point feature vector of the target cylindrical region.

[0107] Following the previous embodiment, in this embodiment, feature encoding of the point cloud sequence data to obtain point feature vectors can be performed as follows: Based on the position information of each point cloud, the point cloud in each frame of the point cloud sequence data is divided into several cylindrical regions (hereinafter referred to as cylinders) at a given resolution, such as 0.5. The cylinder regions can be cuboid cylindrical regions. Due to the sparsity of point clouds, most cylinders are empty, and some cylinders will have very dense point clouds. Therefore, it is necessary to filter the cylinders. For cylinders with more than N point clouds, N points are selected using a random sampling method, and the remaining points are discarded to filter out the target cylinders with N point clouds. Then, the weighted average value (x*, y*, z*) of the point clouds within the cylinder is calculated to augment the features. At the same time, augmentation can be performed by combining the cylinder center coordinates (xc, yc). Specifically, the RANSAC algorithm can be used to cluster the point cloud data in each target cylindrical region into a plane, and feature augmentation can be performed on the point cloud data of the target cylindrical region based on the normal of the plane. The inventors noted that if the region corresponding to a cylinder is a pure ground region (i.e., an unobstructed region), its corresponding normal is close to the z-axis. If the region corresponding to the cylinder contains non-grounded points, the normal of the fitted plane is irregular. For cylinders with a point cloud of 3 or more points, the RANSAC algorithm is used to cluster and fit the points within the cylinder into a plane, and the normal (m, n, p) of this plane is used to augment the point features of the cylinder region. For cylinders with a point cloud of less than three points, since a plane cannot be fitted, (0,0,0) is directly used as the normal of the cylinder for feature augmentation. Finally, the point feature vector of each cylinder after feature augmentation is represented by a twelve-dimensional feature vector (x,y,z,i,xc,yc,x*,y*,z*,m,n,p). It should be noted that in this embodiment, after cylindricalizing the scene point cloud, the resulting point feature vectors are obtained from the cylindrical regions. Local features of these point feature vectors are then extracted and mapped back to the original scene. The resulting scene feature data can be obtained as follows: The PointNet feature extraction module is used to extract features from each cylinder. After cylindricalizing the scene point cloud, the scene can be represented by a tensor of (D, P, N), where D = 12 (i.e., 12-dimensional features), P is the number of all non-empty cylinders in the scene point cloud, and N is the number of points in each cylinder. The PointNet feature extraction network is applied to the D-dimensional features of the N points on these P cylinders. Each of the N point clouds on the P cylinders obtains a C-dimensional feature vector. The entire scene point cloud can be represented by a tensor of (C, P, N), where C is the extracted feature vector of each point cloud, and the meanings of P and N are the same as described above. Then, for each cylinder's N point cloud features, a max pooling operation is applied to obtain its local features, and an output tensor of size (C, P) is constructed, where C is the local features and P is the number of non-empty cylinders.Then, the P feature vectors are mapped onto the original scene according to the xy coordinates at the time of extraction to obtain a pseudo image of size (C, H, W), which represents the scene feature vectors.

[0108] In this embodiment, by cylindricalizing the point cloud in the scene and using the RANSAC algorithm, the point cloud data of the cylindrical region can be fitted into a plane quickly and efficiently, thereby enabling feature augmentation to be completed quickly through the normals of the plane.

[0109] To provide a clearer explanation of the point cloud sequence data processing method provided in this application, the following will be combined with... Figure 6 and one A specific embodiment will be described below, in which a robot is used as an example of the moving object, including the following:

[0110] Step 1: Calibrate the LiDAR, GPS, and IMU so that the GPS, IMU, and LiDAR can synchronously provide the corresponding sensor data and timestamp information.

[0111] Step 2: Acquire point cloud sequence data of the robot's location. Establish global and local coordinate systems for the robot's movement, and adjust the position of the LiDAR to ensure that the LiDAR coordinate system and the robot coordinate system are completely aligned. The LiDAR will provide point cloud data in the local coordinate system; acquire the point cloud sequence data of the robot's location. Fuse GPS and IMU sensor data to obtain the robot's position and orientation information in the global coordinate system.

[0112] Step 3: Point cloud preprocessing. Preprocessing includes coordinate correction and removal of invalid point cloud data.

[0113] Step 4: Point Cloud Feature Encoding. Each frame of the point cloud sequence data is divided into several cylindrical regions. Target cylindrical regions with a sufficient number of points are selected from these regions. The RANSAC algorithm is used to cluster and fit the point cloud data within each target cylindrical region into a plane. Feature augmentation is then performed on the point cloud data of the target cylindrical region based on the plane's normal, resulting in the point feature vectors of the target cylindrical regions.

[0114] Step 5: Point Cloud Feature Extraction. For each cylinder, apply a max-pooling operation to the N point cloud features to obtain its local features, creating an output tensor of size (C, P), where C represents the local features and P is the number of non-empty pillars. Then, map the P feature vectors to the original scene according to the xy coordinates at the time of extraction, obtaining a pseudo-image of size (C, H, W), where H and W represent the length and width of the scene. This pseudo-image can represent the features of the entire scene.

[0115] Step 6: Use the pseudo-image as input to the SegNet network to regress and obtain the elevation feature values ​​of the cylindrical feature vector of the scene (i.e., scene elevation feature values). Based on the scene elevation feature values ​​and the z-values ​​of the point cloud sequence data, locate the obstacle points in the point cloud sequence data.

[0116] Step 7: Perform two-dimensional projection on the obstacle points of each frame point cloud to obtain the occupied grid map sequence data. Each occupied grid map carries time parameters and the position information of the moving object when the occupied grid map is constructed.

[0117] Step 8: Correct the current occupancy grid map. Based on the position information of the moving object in the global coordinate system, the occupancy grid map sequence data is transformed to the global coordinate system. Multiple target occupancy grid maps are extracted from the coordinate-transformed occupancy grid map sequence data. Obstacle clustering is performed on the multiple target occupancy grid maps to obtain the obstacle blocks of each target occupancy grid map. The intersection-union ratio (IUR) of the obstacle blocks in the current occupancy grid map is obtained. If the IUR of each obstacle block in the current occupancy grid map is greater than or equal to the first preset ratio threshold γ, the missing obstacles in the current occupancy grid map are supplemented to correct the current occupancy grid map.

[0118] Step 9: Identify moving obstacle points and label their motion information. If the intersection-union ratio (IUU) of an obstacle block in the occupied grid map at the current moment is less than the first preset ratio threshold γ and further less than the second preset ratio threshold β, then the obstacle block is identified as a moving obstacle point. Obtain the first center coordinates of the moving obstacle point in the occupied grid map at the current moment, and the second center coordinates of the target occupied grid map at the previous moment. Based on the first and second center coordinates, obtain the displacement of the moving obstacle point. Based on the displacement of the moving obstacle point and the time difference between the current moment and the previous moment, determine the motion speed of the moving obstacle point and label the motion speed of the moving obstacle point.

[0119] Step 10: Publish the revised, current-moment occupied grid map, marked with motion obstacle information, to downstream modules to facilitate path planning and navigation.

[0120] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0121] Based on the same inventive concept, this application also provides a point cloud sequence data processing apparatus for implementing the point cloud sequence data processing method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more point cloud sequence data processing apparatus embodiments provided below can be found in the limitations of the point cloud sequence data processing method described above, and will not be repeated here.

[0122] In one embodiment, such as Figure 7 As shown, a point cloud sequence data processing device is provided, including: a data acquisition module 710, an obstacle point localization module 720, an occupied grid map data acquisition module 730, and an occupied grid map correction module 740, wherein:

[0123] The data acquisition module 710 is used to acquire point cloud sequence data of the area where the moving object is located in the target scene.

[0124] The obstacle point localization module 720 is used to locate obstacle points in each frame of the point cloud sequence data.

[0125] The occupancy grid map data acquisition module 730 is used to project obstacle points into each frame point cloud sequence to obtain occupancy grid map sequence data. Each occupancy grid map carries time parameters and the position information of moving objects when constructing the occupancy grid map.

[0126] The occupancy grid map correction module 740 is used to correct the occupancy grid map at the current moment based on time parameters and the position information of moving objects, and to determine the moving obstacle points and their movement information in the occupancy grid map at the current moment.

[0127] The aforementioned point cloud sequence data processing device differs from traditional technologies that use single-frame point cloud data to represent scene information. Instead, it uses multi-frame point cloud data (i.e., point cloud sequence data) as input, locates obstacle points in each frame of the point cloud sequence, and projects these obstacle points to obtain a multi-frame point cloud data occupancy grid map sequence. Then, based on the time parameters and moving object position information in the occupancy grid map, it corrects the current occupancy grid map and determines the moving obstacle points and their motion information. This entire scheme uses multi-frame point cloud data for processing to generate a more accurate occupancy grid map. Furthermore, correcting the current occupancy grid map further improves its accuracy. Additionally, determining the moving obstacle points and their motion information significantly enhances the accuracy of drivable area detection.

[0128] In one embodiment, the position information of the moving object includes the position information of the moving object in the global coordinate system;

[0129] The occupancy grid map correction module 740 is also used to convert the occupancy grid map sequence data to the global coordinate system based on the position information of the moving object in the global coordinate system, extract multiple target occupancy grid maps from the occupancy grid map sequence data after coordinate transformation, the target occupancy grid map is the previous occupancy grid map that overlaps with the occupancy grid map at the current time in the area where the moving object is located, perform obstacle clustering on multiple target occupancy grid maps to obtain the obstacle blocks of each target occupancy grid map, obtain the cross-union ratio of the obstacle blocks in the occupancy grid map at the current time, the cross-union ratio is the overlap rate between the obstacle blocks in the occupancy grid map at the current time and the obstacle blocks in the target occupancy grid map, correct the occupancy grid map at the current time based on the cross-union ratio of the obstacle blocks in the occupancy grid map at the current time, and determine the moving obstacle points and the movement information of the moving obstacle points in the occupancy grid map at the current time.

[0130] In one embodiment, the occupancy grid map correction module 740 is further configured to supplement the missing obstacles in the occupancy grid map at the current time if the intersection-union ratio of each obstacle block in the occupancy grid map at the current time is greater than or equal to a first preset ratio threshold, so as to correct the occupancy grid map at the current time. The missing obstacles are obstacles that are recorded in the target occupancy grid map but not recorded in the occupancy grid map at the current time.

[0131] In one embodiment, the occupancy grid map correction module 740 is further configured to: if the intersection-union ratio of the obstacle block in the occupancy grid map at the current time is less than a second preset ratio threshold, then determine the obstacle block as a moving obstacle point; if the second preset ratio threshold is less than a first preset ratio threshold, obtain the first center coordinates of the moving obstacle point in the occupancy grid map at the current time and the second center coordinates of the target occupancy grid map at the previous time; obtain the displacement of the moving obstacle point based on the first center coordinates and the second center coordinates; and determine the movement speed of the moving obstacle point based on the displacement of the moving obstacle point and the time difference between the current time and the previous time.

[0132] In one embodiment, the obstacle point localization module 720 is further configured to perform feature encoding on the point cloud sequence data to obtain point feature vectors, extract local features from the point feature vectors, map the local features onto the original scene to obtain scene feature data, decode the scene feature data to obtain scene elevation feature values, and locate obstacle points in the point cloud sequence data based on the scene elevation feature values ​​and the coordinate values ​​of the point cloud sequence data.

[0133] In one embodiment, the obstacle point localization module 720 is further configured to divide each frame of point cloud sequence data in the point cloud sequence data into several cylindrical regions, select target cylindrical regions from the several cylindrical regions whose point cloud quantity meets the preset quantity requirement, use the RANSAC algorithm to cluster and fit the point cloud data in each target cylindrical region into a plane, and perform feature augmentation on the point cloud data of the target cylindrical region according to the normal of the plane to obtain the point feature vector of the target cylindrical region.

[0134] Each module in the aforementioned point cloud sequence data processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0135] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores point cloud sequence data and occupancy raster maps, among other data. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a point cloud sequence data processing method.

[0136] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0137] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the point cloud sequence data processing method described above.

[0138] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the point cloud sequence data processing method described above.

[0139] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the point cloud sequence data processing method described above.

[0140] It should be noted that the data involved in this application (including but not limited to data used for analysis, data stored, data displayed, etc.) are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0141] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0142] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0143] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for processing point cloud sequence data, characterized in that, The method includes: Acquire point cloud sequence data of the area where the moving object is located in the target scene; Locate the obstacle points in each frame of the point cloud sequence data; The obstacle points of each frame point cloud sequence are projected to obtain the occupied grid map sequence data. Each occupied grid map carries time parameters and the position information of the moving object when the occupied grid map is constructed. The position information of the moving object includes the position information of the moving object in the global coordinate system. Based on the position information of the moving object in the global coordinate system, the occupied grid map sequence data is converted to the global coordinate system; Multiple target occupancy grid maps are extracted from the occupancy grid map sequence data after coordinate transformation. The target occupancy grid map is a previous occupancy grid map that overlaps with the occupancy grid map at the current moment in the area where the moving object is located. Obstacle clustering is performed on the multiple target-occupied grid maps to obtain obstacle blocks for each target-occupied grid map; Obtain the intersection-over-union ratio (IoU) of obstacle blocks in the occupied grid map at the current time, where IoU is the overlap rate between the obstacle blocks in the occupied grid map at the current time and the obstacle blocks in the target occupied grid map. Based on the intersection-union ratio of obstacle blocks in the occupied grid map at the current moment, the occupied grid map at the current moment is corrected, and the moving obstacle points in the occupied grid map at the current moment and the movement information of the moving obstacle points are determined.

2. The method according to claim 1, characterized in that, The step of correcting the occupancy grid map at the current moment based on the intersection-union ratio of each obstacle block in the occupancy grid map at the current moment includes: If the intersection-union ratio of each obstacle block in the current occupied grid map is greater than or equal to the first preset ratio threshold, then the missing obstacles in the current occupied grid map are supplemented to correct the current occupied grid map. The missing obstacles are obstacles recorded in the target occupied grid map but not recorded in the current occupied grid map.

3. The method according to claim 2, characterized in that, Based on the intersection-over-union ratio of each obstacle block in the occupied grid map at the current moment, the moving obstacle points in the occupied grid map at the current moment and the motion information of the moving obstacle points are determined, including: If the intersection-union ratio of the obstacle block occupying the grid map at the current moment is less than the second preset ratio threshold, then the obstacle block is identified as a moving obstacle point, and the second preset ratio threshold is less than the first preset ratio threshold. Obtain the first center coordinates of the occupied grid map of the moving obstacle point at the current time, and the second center coordinates of the occupied grid map of the target at the previous time; The displacement of the obstacle point is obtained based on the first center coordinate and the second center coordinate; The speed of the moving obstacle is determined based on its displacement and the time difference between the current moment and the previous moment.

4. The method according to any one of claims 1 to 3, characterized in that, The process of locating obstacle points in each frame of the point cloud sequence data includes: Each frame of the point cloud sequence in the point cloud sequence data is feature-encoded to obtain a point feature vector; Local features are extracted from the point feature vectors and mapped onto the original scene to obtain scene feature data; Decode the scene feature data to obtain the scene elevation feature value; Based on the scene elevation feature value and the coordinate value of the point cloud sequence data, the obstacle point in each frame of the point cloud sequence data is located.

5. The method according to claim 4, characterized in that, The step of performing feature encoding on each frame of the point cloud sequence data to obtain a point feature vector includes: Each frame of point cloud sequence data in the point cloud sequence data is divided into several cylindrical regions; Select target cylindrical regions from the plurality of cylindrical regions whose point cloud quantity meets the preset quantity requirement; The RANSAC algorithm is used to cluster the point cloud data in each target cylindrical region into a plane. The point cloud data of the target cylindrical region is then augmented with features based on the normal of the plane to obtain the point feature vector of the target cylindrical region.

6. A point cloud sequence data processing device, characterized in that, The device includes: The data acquisition module is used to acquire point cloud sequence data of the area where moving objects are located in the target scene; An obstacle point localization module is used to locate obstacle points in each frame of the point cloud sequence data. The occupancy grid map data acquisition module is used to project the obstacle points of each frame point cloud sequence to obtain occupancy grid map sequence data. Each occupancy grid map carries time parameters and the position information of the moving object when constructing the occupancy grid map. The position information of the moving object includes the position information of the moving object in the global coordinate system. The occupancy grid map correction module is used to convert the occupancy grid map sequence data to the global coordinate system based on the position information of the moving object in the global coordinate system, extract multiple target occupancy grid maps from the occupancy grid map sequence data after coordinate transformation, wherein the target occupancy grid map is a previous occupancy grid map that overlaps with the occupancy grid map at the current time in the area where the moving object is located, perform obstacle clustering on the multiple target occupancy grid maps to obtain obstacle blocks of each target occupancy grid map, obtain the intersection-union ratio (IUGR) of the obstacle blocks in the occupancy grid map at the current time, wherein the IUGR is the overlap rate between the obstacle blocks in the occupancy grid map at the current time and the obstacle blocks in the target occupancy grid maps, correct the occupancy grid map at the current time based on the IUGR of the obstacle blocks in the occupancy grid map at the current time, and determine the moving obstacle points in the occupancy grid map at the current time and the motion information of the moving obstacle points.

7. The apparatus according to claim 6, characterized in that, The occupancy grid map correction module is further configured to supplement the missing obstacles in the occupancy grid map at the current time if the intersection-union ratio of each obstacle block in the occupancy grid map at the current time is greater than or equal to a first preset ratio threshold, so as to correct the occupancy grid map at the current time. The missing obstacles are obstacles that are recorded in the target occupancy grid map but not recorded in the occupancy grid map at the current time.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images