Point cloud processing method, apparatus, and electronic device

Abstract

Embodiments of the present application disclose a point cloud processing method and device and electronic equipment. The method comprises: obtaining a plurality of to-be-processed data points in an original point cloud; obtaining a plurality of high-reflection areas and a plurality of two-dimensional feature maps based on the plurality of to-be-processed data points; inputting the plurality of two-dimensional feature maps into a pre-trained deep learning model to obtain boundary information corresponding to each of the plurality of high-reflection areas; obtaining a target area in each high-reflection area based on shape category information and position information corresponding to each of the plurality of high-reflection areas; and filtering out data points in the plurality of high-reflection areas that are not in the corresponding target areas from the original point cloud to obtain a target point cloud. In this way, the data points in the original point cloud except for the target areas can be filtered out to obtain the target point cloud, thereby improving the accuracy of filtering out noise points on the surface of a target object.

Description

Technical Field

[0001] This application relates to the field of point cloud processing technology, and more specifically, to a point cloud processing method, apparatus, and electronic device. Background Technology

[0002] With the widespread application of LiDAR technology, processing the point cloud data acquired by LiDAR has become a research hotspot. For example, filtering out point cloud noise is crucial. Point cloud noise refers to invalid points acquired by LiDAR, which can easily cause false detections in target detection algorithms based on LiDAR point clouds and registration failures in high-precision point cloud matching algorithms. In practical applications, point cloud noise can be caused by surface noise of the target object, such as the properties of the target object's surface material (high reflectivity surfaces cause excessive energy reflection in the point cloud, making the target object appear larger than it actually is) and roughness (uneven surfaces cause changes in the emission angle of the point cloud). Correlation methods can filter out point cloud noise through plane fitting, but the accuracy of these methods still needs improvement. Summary of the Invention

[0003] In view of the above problems, this application proposes a point cloud processing method, apparatus, and electronic device to improve the above problems.

[0004] In a first aspect, this application provides a point cloud processing method, the method comprising: acquiring multiple data points to be processed from an original point cloud, wherein the data points to be processed are data points with reflectance greater than a reflectance threshold; obtaining multiple high-reflectance regions and multiple two-dimensional feature maps based on the multiple data points to be processed, wherein the multiple high-reflectance regions and the multiple two-dimensional feature maps are... Figure 1 In a one-to-one correspondence, each high-reflectivity region is a region composed of a set of multiple data points to be processed located within a specified circle, and each two-dimensional feature map is obtained by projecting the corresponding high-reflectivity region onto a two-dimensional plane; the multiple two-dimensional feature maps are input into a pre-trained deep learning model to obtain the boundary information corresponding to each of the multiple high-reflectivity regions, the boundary information including shape category information and position information; based on the shape category information and position information corresponding to each of the multiple high-reflectivity regions, the target region in each of the high-reflectivity regions is obtained, the target region being the region where the target object is located in the original point cloud; data points in the multiple high-reflectivity regions that are not in the corresponding target region are filtered out from the original point cloud to obtain the target point cloud.

[0005] Secondly, this application provides a point cloud processing apparatus, the apparatus comprising: a data acquisition unit, configured to acquire an original point cloud, the original point cloud including a plurality of data points, each of the plurality of data points corresponding to a reflectance; acquire a plurality of data points to be processed, the data points to be processed being data points whose corresponding reflectance is greater than a reflectance threshold; and a high-reflectance region acquisition unit, configured to obtain a plurality of high-reflectance regions and a plurality of two-dimensional feature maps based on the plurality of data points to be processed, wherein the plurality of high-reflectance regions and the plurality of two-dimensional feature maps are... Figure 1 In a one-to-one correspondence, each high-reflectivity region is a region composed of a set of multiple data points to be processed located within a specified circle, and each two-dimensional feature map is obtained by projecting the corresponding high-reflectivity region onto a two-dimensional plane; the target region acquisition unit is used to input the multiple two-dimensional feature maps into a pre-trained deep learning model to obtain the boundary information corresponding to each of the multiple high-reflectivity regions, the boundary information including shape category information and position information; based on the shape category information and position information corresponding to each of the multiple high-reflectivity regions, the target region in each of the high-reflectivity regions is obtained, the target region being the region where the target object is located in the original point cloud; the target point cloud acquisition unit is used to filter out the data points in the multiple high-reflectivity regions that are not in the corresponding target region from the original point cloud to obtain the target point cloud.

[0006] Thirdly, this application provides an electronic device including one or more processors and a memory; one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs being configured to perform the methods described above.

[0007] Fourthly, this application provides a computer-readable storage medium storing program code, wherein the above-described method is executed when the program code is run.

[0008] This application provides a point cloud processing method, apparatus, electronic device, and storage medium. After acquiring multiple data points to be processed in the original point cloud with reflectivity greater than a reflectivity threshold, multiple high-reflectivity regions composed of sets of multiple data points to be processed within a specified circle are obtained based on the multiple data points to be processed, and multiple two-dimensional feature maps obtained by projecting the corresponding high-reflectivity regions onto a two-dimensional plane are obtained. The multiple two-dimensional feature maps are input into a pre-trained deep learning model to obtain boundary information including shape category information and position information for each of the multiple high-reflectivity regions. Based on the shape category information and position information corresponding to each of the multiple high-reflectivity regions, target regions containing the target objects in the original point cloud are obtained for each of the multiple high-reflectivity regions. Data points in the multiple high-reflectivity regions that are not in the corresponding target regions are filtered out from the original point cloud to obtain the target point cloud. The above method enables the generation of multiple high-reflectivity regions and their corresponding two-dimensional feature maps based on multiple data points to be processed in the original point cloud. Based on the two-dimensional feature maps and a pre-trained deep learning model, the shape category information and location information of each of the multiple high-reflectivity regions are obtained. Then, based on the shape category information and location information of each of the multiple high-reflectivity regions, the region where the target object is located in the original point cloud is obtained. Thus, by filtering out data points other than the target region in the original point cloud, the target point cloud can be obtained, thereby improving the accuracy of noise filtering on the surface of the target object. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A flowchart of a point cloud processing method proposed in an embodiment of this application is shown;

[0011] Figure 2 A schematic diagram of a normal raw point cloud as proposed in this application is shown;

[0012] Figure 3 A schematic diagram of a raw point cloud exhibiting high anti-inflation phenomenon, as proposed in this application, is shown.

[0013] Figure 4 This application shows Figure 1 A flowchart of one implementation method of S120;

[0014] Figure 5 A schematic diagram of a target region shape type proposed in this application is shown;

[0015] Figure 6A schematic diagram of a deep learning model proposed in this application is shown;

[0016] Figure 7 A schematic diagram of a target area proposed in this application is shown;

[0017] Figure 8 This paper presents a schematic diagram of a method for filtering data points in high-reflectivity regions according to the present application.

[0018] Figure 9 A flowchart of a point cloud processing method according to another embodiment of this application is shown;

[0019] Figure 10 This paper shows a structural block diagram of a point cloud processing device according to an embodiment of the present application;

[0020] Figure 11 A structural block diagram of an electronic device proposed in this application is shown. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0022] In this application embodiment, the inventors propose a point cloud processing method, apparatus, and electronic device. After acquiring multiple data points to be processed in the original point cloud with reflectivity greater than a reflectivity threshold, multiple high-reflectivity regions composed of sets of multiple data points to be processed within a specified circle are obtained based on the multiple data points to be processed, and multiple two-dimensional feature maps obtained by projecting the corresponding high-reflectivity regions onto a two-dimensional plane are obtained. The multiple two-dimensional feature maps are input into a pre-trained deep learning model to obtain boundary information including shape category information and position information corresponding to each of the multiple high-reflectivity regions. Based on the shape category information and position information corresponding to each of the multiple high-reflectivity regions, target regions containing the target objects in the original point cloud are obtained in each of the multiple high-reflectivity regions. Data points in the multiple high-reflectivity regions that are not in the corresponding target regions are filtered out from the original point cloud to obtain the target point cloud. The above method enables the generation of multiple high-reflectivity regions and their corresponding two-dimensional feature maps based on multiple data points to be processed in the original point cloud. Based on the two-dimensional feature maps and a pre-trained deep learning model, the shape category information and location information of each of the multiple high-reflectivity regions are obtained. Then, based on the shape category information and location information of each of the multiple high-reflectivity regions, the region where the target object is located in the original point cloud is obtained. Thus, by filtering out data points other than the target region in the original point cloud, the target point cloud can be obtained, thereby improving the accuracy of noise filtering on the surface of the target object.

[0023] Please see Figure 1 This application provides a point cloud processing method, the method comprising:

[0024] S110: Obtain multiple data points to be processed from the original point cloud, wherein the data points to be processed are data points with reflectivity greater than the reflectivity threshold.

[0025] Here, the original point cloud refers to the point cloud data scanned by the LiDAR. One frame of point cloud data corresponds to one frame of original point cloud. Each frame of point cloud data can contain multiple data points. Each data point can be a data point to be processed or a non-high reflectivity point. A non-high reflectivity point refers to a data point with a reflectivity less than or equal to a reflectivity threshold. Multiple data points to be processed and multiple non-high reflectivity points can each correspond to three-dimensional coordinates and reflectivity (intensity). For example, a data point A to be processed can be represented as (X... A Y A Z A Int A Optionally, the reflectivity threshold can be obtained based on multiple experiments.

[0026] Taking LiDAR scanning in a vehicle scenario as an example, when the LiDAR scans the surface of some targets (such as road signs), noise may appear due to the high energy (high reflectivity) of the point cloud reflected back by the target. This noise may cause the target to appear as high-reflectivity dilatation in the original point cloud. High-reflectivity dilatation refers to the phenomenon where, after the LiDAR scans the surface of a high-reflectivity target, the corresponding point cloud data of the target appears to diffuse outwards in the original point cloud, making the original target point cloud image appear larger. For example... Figure 2 This shows the correct point cloud image of the target object (road sign) in the original point cloud. Figure 3 The image shows an inflated point cloud image of a target object (road sign) in the original point cloud due to high-reflection dilation. When high-reflection ghosting or high-reflection dilation occurs in the original point cloud, it may lead to problems such as false detection of obstacles and unstable obstacle detection in the LiDAR-based environmental perception algorithm. This can further cause abnormal actions such as false braking in the downstream planning and control algorithm of the environmental perception algorithm, affecting the safety and comfort of the vehicle.

[0027] One approach is to acquire the raw point cloud scanned by lidar and use the data points in the raw point cloud with reflectivity greater than the reflectivity threshold as data points to be processed, thereby obtaining multiple data points to be processed in the raw point cloud.

[0028] It should be noted that the execution subject of the point cloud processing method proposed in this application can be a vehicle or an electronic device, and the electronic device can refer to a cloud server, etc. When the execution subject is a vehicle, the original point cloud can be acquired in real time through LiDAR. When the execution subject is an electronic device, the vehicle can collect and store the original point cloud in real time, and periodically send the stored original point cloud to the electronic device so that the electronic device can obtain the original point cloud.

[0029] In this embodiment of the application, non-high reflectivity points in the original point cloud are filtered out by using a reflectivity threshold, which narrows the search range for the target area and improves the efficiency of acquiring the target area.

[0030] S120: Based on the multiple data points to be processed, multiple high-reflectivity regions and multiple two-dimensional feature maps are obtained, wherein the multiple high-reflectivity regions and the multiple two-dimensional feature maps are... Figure 1 In a one-to-one correspondence, each of the high-reflectivity regions is a region consisting of a set of multiple data points to be processed located in a specified circle, and each of the two-dimensional feature maps is obtained by projecting the corresponding high-reflectivity region onto a two-dimensional plane.

[0031] One approach is to cluster multiple data points to be processed to obtain multiple high-reflection regions.

[0032] The center of the designated circle for each high-reflectivity region can be the center point of the corresponding high-reflectivity region.

[0033] Clustering divides a dataset into different classes or clusters based on a specific criterion (such as distance), maximizing the similarity of data objects within the same cluster while maximizing the differences between data objects in different clusters. Each cluster can correspond to a cluster center, which can be the center point of a high-intensity region.

[0034] As one approach, during the current processing, one data point can be selected from multiple data points to be processed as the initial center point. Based on the data points to be processed located within a specified circle centered on the initial center point, an initial set is obtained, where the radius of the specified circle can be the initial radius. All data points to be processed in the initial set are clustered to obtain high-inflection regions in the current processing. All data points corresponding to the high-inflection regions in the current processing are deleted from the multiple data points to obtain the remaining data points to be processed in the current processing. These remaining data points are then used as multiple data points to be processed in the next processing, until no data points remain, thus obtaining multiple high-inflection regions.

[0035] Optionally, one data point can be randomly selected from multiple data points to be processed as the initial center point.

[0036] Optionally, points other than the initial center point among multiple data points to be processed can be defined as search points. Whether a point belongs to the initial set can be determined based on the distances of each search point to the initial center point. If the distance between a search point and the initial center point is less than or equal to the initial radius, the search point belongs to the initial set; if the distance is greater than the initial radius, the search point does not belong to the initial set. The distances between each search point and the initial center point can be Euclidean distance, Manhattan distance, etc. The formula for calculating Euclidean distance is as follows:

[0037]

[0038] The initial center point can be (X0, Y0, Z0), and the search point can be (X1, Y1, Z1).

[0039] For example, multiple data points to be processed can be {p1, p2, p3, ..., pn}, the initial radius can be r = 0.5 meters, the initial center point can be p1, the three-dimensional coordinates of p1 can be (0, 0, 0), the search points can be {p2, p3, ..., pn}, the three-dimensional coordinates of p2 can be (0.2, 0.1, 0), and the three-dimensional coordinates of p3 can be (0.9, 0.1, 0). Then the distance between p2 and p1 can be 0.22 meters, and the distance between p3 and p1 can be 0.95 meters. Therefore, it can be determined that p2 belongs to the initial set, and p3 does not belong to the initial set.

[0040] Optional, such as Figure 4 As shown, the step of clustering all data points to be processed in the initial set to obtain high-inflection regions in the current processing includes:

[0041] S121: Perform coarse clustering based on the initial center point and initial radius corresponding to the initial set to obtain the reference center point.

[0042] Among them, coarse clusters are more likely to meet the clustering process termination condition than fine clusters.

[0043] In one approach, during the current coarse clustering process, the updated center point can be obtained based on the average of the three-dimensional coordinates of all data points to be processed within the initial set. If the distance between the initial center point and the updated center point in the current coarse clustering process is greater than or equal to a first distance threshold, the updated center point in the current coarse clustering process can be used as the initial center point in the next coarse clustering process. If the distance between the initial center point and the updated center point in the current coarse clustering process is less than the first distance threshold, the updated center point in the current coarse clustering process can be used as the reference center point in the current coarse clustering process.

[0044] Optionally, the average of the three-dimensional coordinates of all data points to be processed in the initial set can be used as the update center point in the current coarse clustering process if the average value belongs to the initial set; otherwise, the data point in the initial set that is closest to the average value can be used as the update center point in the current coarse clustering process.

[0045] Optionally, the average coordinates of each dimension of all data points to be processed within the initial set can be calculated separately to obtain the average of the three-dimensional coordinates corresponding to each data point within the initial set. The formula for calculating the average can be:

[0046]

[0047] Where m can represent the total number of data points to be processed in the initial set, (X i ,Y i Z i ) can represent the three-dimensional coordinates of the i-th data point to be processed in the initial set.

[0048] Optionally, the initial radius and the first distance threshold can be determined based on the point resolution and angular resolution of the lidar.

[0049] Optionally, the distance between the initial centroid and the updated centroid in the coarse clustering process can be Euclidean distance, Manhattan distance, etc.

[0050] For example, the first distance threshold can be th1 = 0.1 meters, the initial radius can be r = 0.5 meters, the initial center can be p1 in the current coarse clustering process, the three-dimensional coordinates of p1 can be (0, 0, 0), the updated center point in the current coarse clustering process can be p2, the three-dimensional coordinates of p2 can be (0.2, 0.1, 0), the distance between p2 and p1 can be 0.22 meters, so p2 can be used as the initial center point in the next coarse clustering process, and based on p2 and the initial radius, the initial set in the next coarse clustering process can be obtained, and based on the average of the three-dimensional coordinates of all the data points to be processed in the initial set in the next coarse clustering process, the updated center point in the next coarse clustering process can be p5, the three-dimensional coordinates of p5 can be (0.15, 0.1, 0), the distance between p5 and p2 can be 0.05 meters, so p5 can be used as the reference center point.

[0051] S122: Based on the preset step size and the initial radius, obtain the updated radius.

[0052] The preset step size can refer to the pre-determined update range, which can be determined based on the point resolution and angular resolution of the lidar.

[0053] One approach is to use the sum of the preset step size and the initial radius as the updated radius.

[0054] For example, if the preset step size is 0.1 meters and the initial radius is 0.5 meters, then the update radius can be 0.6 meters.

[0055] S123: Based on the data points to be processed located within a specified circle centered on the reference center point, an update set is obtained, and the reference center point is used as the initial center point of the update set, and the update radius is used as the initial radius of the update set, wherein the radius of the specified circle is the update radius.

[0056] In one approach, an update set can be obtained based on the data points to be processed located within a specified circle centered on the reference center point obtained in step S121, wherein the radius of the specified circle is the update radius obtained in step S122.

[0057] S124: Perform fine clustering based on the initial center point and initial radius corresponding to the updated set to obtain the high-inflection region in the current processing.

[0058] As one approach, during the current fine clustering process, the updated set can be used as the initial set for that process. Based on the average of the three-dimensional coordinates of all data points to be processed in the initial set, the updated center point is obtained. Based on the data points to be processed located within a specified circle centered on the updated center point, an updated set is obtained; the radius of this specified circle can be the initial radius corresponding to the initial set. If the distance between the initial center point and the updated center point is greater than or equal to a second distance threshold... Alternatively, if the change in the number of data points to be processed in the initial set and the updated set during the current fine clustering process is greater than or equal to the number of points threshold, coarse clustering is performed based on the updated centroid, updated set, and initial radius during the current fine clustering process to obtain the initial set, initial centroid, and initial radius for the next fine clustering process. If the distance between the initial centroid and the updated centroid during the current fine clustering process is less than the second distance threshold and the change in the number of data points to be processed in the initial set and the updated set during the current fine clustering process is less than the number of points threshold, the updated reference centroid is used as the centroid, and the high-reflection region during the current processing is obtained based on the updated set.

[0059] Optionally, the second distance threshold can be greater than the first distance threshold, and the second distance threshold and the point count threshold can be determined based on the point resolution and angular resolution of the LiDAR. For example, the first distance threshold can be 0.1 meters, the second distance threshold can be 0.2 meters, and the point count threshold can be 3.

[0060] For example, in the current fine clustering process, the second distance threshold can be th2 = 0.2 meters, the number of points threshold can be 3, the initial radius can be r = 0.6 meters, the initial center in the current fine clustering process can be P5, and the three-dimensional coordinates of p5 can be (0.15, 0.1, 0). Based on P5 and the initial radius, the initial set in the current fine clustering process can be obtained, and the total number of all data points to be processed in the initial set can be 25. Based on the average of the three-dimensional coordinates of all the data points to be processed in the initial set during the current fine clustering process, the updated center point in the current fine clustering process can be P8, and the three-dimensional coordinates of p8 can be (0.15, 0.25, 0). The distance between p5 and p8 can be 0.15 meters. Since the distance between p5 and p8 is less than the second distance threshold, the updated set in the current fine clustering process can be obtained based on P8 and the initial radius in the current fine clustering process. The total number of data points to be processed in the updated set can be 30. The change in the number of data points to be processed between the initial set and the updated set in the current fine clustering process is 5 (30-25). This change is greater than the number of points threshold. Therefore, coarse clustering needs to be performed based on the updated center point, the updated set, and the initial radius in the current fine clustering process to obtain the initial set, the initial center point, and the initial radius for the next fine clustering process.

[0061] When performing coarse clustering based on the updated centroids, updated set, and initial radius in the current fine clustering process, the corresponding first distance threshold can be th1 = 0.1 meters, the initial radius can be r = 0.6 meters, and the initial centroid can be p8. Then, based on the method in step S121, coarse clustering can be performed on the updated centroids, updated set, and initial radius in the current fine clustering process, resulting in a reference centroid P10 with three-dimensional coordinates (0.2, 0.25, 0). The initial centroid for the next fine clustering process can then be P10, the initial radius can be 0.7(0.6+0.1) meters, the second distance threshold can be th2 = 0.2 meters, and the point count threshold can be 3. Based on P10 and the initial radius, the initial set for the next fine clustering process can be obtained, and the total number of all unprocessed data points in this initial set can be 28.

[0062] Based on the average of the three-dimensional coordinates of all the data points to be processed in the initial set during the next fine clustering process, the update center point in the current fine clustering process can be P12, and the three-dimensional coordinates of p12 can be (0.3, 0.25, 0). The distance between p12 and p10 can be 0.1 meters. Since the distance between p12 and p10 is less than the second distance threshold, the update set {p8, p10, p12, ...} in the current fine clustering process can be obtained based on P12 and the initial radius. The total number of data points to be processed in this update set can be 30. The change in the number of data points to be processed in the initial set and the update set in the current fine clustering process is 2(30-28). This change is less than the number of points threshold. Therefore, p12 can be used as the center point, and the high-reflection region in the current processing process can be obtained based on the update set.

[0063] In this embodiment of the application, multiple high-reflection regions are obtained by clustering multiple data points to be processed, which can further narrow the search range of the target region and improve the efficiency of locating the target region.

[0064] One approach is to obtain multiple high-reflectivity regions, and then perform planar fitting on each of the multiple high-reflectivity regions based on their respective center points to generate two-dimensional feature maps for each of the multiple high-reflectivity regions.

[0065] In this context, a two-dimensional feature map refers to a feature map in an image coordinate system. The size of a two-dimensional feature map can be represented as (W, H). If the resolution of a single point in the original point cloud is sp, then the two-dimensional feature map can represent a point cloud region of (W*sp)*(H*sp). For example, if sp can be 0.2, W can be 128, and H can be 128, then a 128*128 two-dimensional feature map can be generated. This 128*128 two-dimensional feature map can represent a point cloud projection region of 25.6m*25.6m, and each pixel in the two-dimensional feature map can represent a 0.2m*0.2m region.

[0066] One approach is to perform planar fitting on multiple high-reflectivity regions based on their respective center points and the least squares method, in order to generate two-dimensional feature maps corresponding to each of the multiple high-reflectivity regions.

[0067] S130: Input the multiple two-dimensional feature maps into a pre-trained deep learning model to obtain the boundary information corresponding to each of the multiple high-reflectivity regions. The boundary information includes shape category information and position information.

[0068] Among them, the shape category information corresponding to the high-reflection region can be understood as the shape category information of the area where the target object is located in the two-dimensional feature map corresponding to the high-reflection region, and the location information corresponding to the high-reflection region can be understood as the location information of the area where the target object is located in the two-dimensional feature map corresponding to the high-reflection region.

[0069] Here, a deep learning model can refer to a multi-task deep learning model. A deep learning model can be used to perform a classification task, which can be used to determine the shape type of the area where the target object is located. The shape category information can include the predicted probabilities of various shapes.

[0070] Deep learning models can also be used to perform regression tasks, which can output location information of the area where the target object is located. The location information can include the coordinates of the pixel and the range of the pointer to the pixel.

[0071] like Figure 5 As shown, the shape of the region containing the target object can be a first rectangle, a second rectangle, a third rectangle, a circle, or a triangle. Each shape can be represented by three directed line segments, with an angle of 120 degrees between any two directed line segments. The vertically upward direction can be defined as the 0-degree orientation, and clockwise rotation as the positive direction of rotation. The 120-degree and 240-degree directed line segments are of equal length. Based on these constraints, each region containing the target object can be represented as (type, ix, iy, l1, l2, l3). Here, type represents the shape of the region containing the target object; ix and iy are the pixel coordinates of the starting points of the vectors corresponding to the three directed line segments in the two-dimensional feature map; l1, l2, and l3 represent the normalized vector lengths, with values ​​ranging from -1 to 1. The positional information can include the pixel coordinates of the starting points of the vectors corresponding to the three directed line segments in the two-dimensional feature map and the vector length of each directed line segment.

[0072] One approach is to input multiple two-dimensional feature maps into a pre-trained deep learning model to obtain the boundary information corresponding to each of the multiple high-reflectivity regions.

[0073] For example, a pre-trained deep learning model can be as follows: Figure 6As shown, after inputting the two-dimensional feature maps corresponding to multiple high-contrast regions into a pre-trained deep learning model, a 6-dimensional classification vector representing the shape type of the region where the target object is located and a 5-dimensional regression vector representing the location information can be obtained for each two-dimensional feature map. The first dimension of the 6-dimensional classification vector represents whether the target object exists in the two-dimensional feature map, and the second to sixth dimensions represent the probabilities that the target object's region in the two-dimensional feature map is a first rectangle, a second rectangle, a third rectangle, a circle, and a triangle, respectively. The 5-dimensional regression vector represents the pixel coordinates of the starting points of the vectors corresponding to the three directed line segments in the two-dimensional feature map and the vector length of each directed line segment. Based on the 6-dimensional classification vector and the 5-dimensional regression vector, the following can be obtained: Figure 7 The area where the target object is located is shown.

[0074] S140: Based on the shape category information and location information corresponding to each of the multiple high-reflectivity regions, a target region is obtained in each of the high-reflectivity regions, and the target region is the region where the target object is located in the original point cloud.

[0075] One approach is to obtain a reference high-reflection region for each high-reflection region based on the shape category information and location information corresponding to each of the multiple high-reflection regions. The reference high-reflection region is the region where the target object is located in the corresponding two-dimensional feature map. The reference high-reflection region corresponding to each high-reflection region is then back-projected onto the original point cloud to obtain the target region in each high-reflection region.

[0076] Optionally, the shape of each reference high-reflection region can be obtained based on the predicted probabilities of various shapes corresponding to each high-reflection region; the range of each reference high-reflection region can be obtained based on the pointing pixel coordinates and the indicating range of the pointing pixel coordinates corresponding to each high-reflection region; and the reference high-reflection region corresponding to each high-reflection region can be obtained based on the shape and range of the reference high-reflection region.

[0077] S150: Filter out the data points in the multiple high-reflectivity regions that are not in the corresponding target regions from the original point cloud to obtain the target point cloud.

[0078] In this embodiment of the application, since the main purpose is to accurately locate the position of the target object where the high-reflection expansion phenomenon occurs, the data points in the high-reflection region that are not in the corresponding target region can be called expansion points.

[0079] As a way, such as Figure 8 As shown, the coordinates of all data points in the target area corresponding to each of the multiple high-contrast regions can be obtained. Based on the obtained coordinates, data points in the corresponding high-contrast regions that are not in the corresponding target area can be filtered out to obtain the target point cloud.

[0080] This embodiment provides a point cloud processing method. After acquiring multiple data points in the original point cloud with reflectivity greater than a reflectivity threshold, multiple high-reflectivity regions composed of sets of multiple data points in a specified circle are obtained based on the multiple data points to be processed, and multiple two-dimensional feature maps obtained by projecting the corresponding high-reflectivity regions onto a two-dimensional plane are obtained. The multiple two-dimensional feature maps are input into a pre-trained deep learning model to obtain boundary information including shape category information and position information for each of the multiple high-reflectivity regions. Based on the shape category information and position information corresponding to each of the multiple high-reflectivity regions, target regions containing the target objects in the original point cloud are obtained for each of the multiple high-reflectivity regions. Data points in the multiple high-reflectivity regions that are not in the corresponding target regions are filtered out from the original point cloud to obtain the target point cloud. The above method enables the generation of multiple high-reflectivity regions and their corresponding two-dimensional feature maps based on multiple data points to be processed in the original point cloud. Based on the two-dimensional feature maps and a pre-trained deep learning model, the shape category information and location information of each of the multiple high-reflectivity regions are obtained. Then, based on the shape category information and location information of each of the multiple high-reflectivity regions, the region where the target object is located in the original point cloud is obtained. Thus, by filtering out data points other than the target region in the original point cloud, the target point cloud can be obtained, thereby improving the accuracy of noise filtering on the surface of the target object.

[0081] Please see Figure 9 This application provides a point cloud processing method, the method comprising:

[0082] S210: Obtain multiple data points to be processed from the original point cloud, wherein the data points to be processed are data points with reflectivity greater than the reflectivity threshold.

[0083] S220: Cluster the multiple data points to be processed to obtain multiple high-reflection regions. Each high-reflection region is a region composed of a set of multiple data points to be processed located in a specified circle, and the center of the specified circle is the center point of the high-reflection region.

[0084] S230: Fit all the data points to be processed in each of the high-reflectivity regions using the least squares method to obtain the plane and target data points corresponding to each of the high-reflectivity regions.

[0085] As one approach, during the processing of a high-reflectivity region, all data points to be processed in the high-reflectivity region can be fitted using the least squares method to obtain a reference plane corresponding to the high-reflectivity region. If the angle between the reference plane and the specified plane is greater than an angle threshold, multiple third distances can be obtained. These multiple third distances can be the distances between each data point to be processed in the high-reflectivity region and the reference plane. Interference data points are filtered out from all data points to be processed to obtain filtered data points. The least squares method is then used to fit the filtered data points until the angle between the reference plane and the specified plane corresponding to the high-reflectivity region is less than or equal to the angle threshold. The reference plane with an angle less than or equal to the angle threshold is taken as the plane of the high-reflectivity region. Interference data points are the data points ranked first in the specified position after sorting all data points to be processed in the high-reflectivity region according to their respective third distances in descending order.

[0086] Optionally, the step of fitting all the data points to be processed in the high-reflectivity region using the least squares method to obtain the reference plane corresponding to the high-reflectivity region may include: first assuming that the equation of the reference plane is Ax + By + Cz + D = 0, and the objective function is d = ax + by - z, then the problem of plane fitting of the high-reflectivity region based on the least squares plane can be transformed into minimizing the objective function d. The objective function d can satisfy that the partial derivatives with respect to the three variables x, y, and z are 0, and thus a system of equations can be constructed for solving. The formula for calculating d is as follows;

[0087]

[0088] Where n can represent the total number of all data points to be processed in the high-reflection region, (x i ,y i ,z i Let ) represent the three-dimensional coordinates of the i-th data point to be processed in the high-reflectivity region, and A, B, and C represent the coefficients of the reference plane. Further, by constructing constraint equations to solve for the coefficients of the reference plane, the coefficients of the reference plane can be obtained, and thus the fitting equation of the reference plane can be obtained. The constraint equations are as follows:

[0089]

[0090] In this embodiment, the designated plane can refer to the ground. The included angle threshold can be determined based on actual needs. For example, in this application, the target objects are mostly road signs, while in actual roads, road signs are mostly perpendicular to the ground, so the included angle threshold can be 120 degrees. The preceding position can be obtained based on multiple trials, such as the first 5 digits.

[0091] In this embodiment, by filtering out interference data points in the high-reflectivity region, the angle between the reference plane corresponding to the high-reflectivity region and the specified plane can be less than or equal to the angle threshold, thereby making the obtained reference plane close to the fitting plane corresponding to the target region, improving the accuracy of the obtained target region, and thus improving the accuracy of downstream tasks (such as control algorithms).

[0092] S240: Obtain a first distance and a second distance corresponding to each of the high-reflection regions, wherein the first distance is the distance between each of the plurality of non-high-reflection points and the plane corresponding to each of the high-reflection regions, and the second distance is the minimum distance between each of the plurality of non-high-reflection points and all data points to be processed in each of the high-reflection regions.

[0093] One approach is to obtain the first and second distances corresponding to each high-reflection region based on the three-dimensional coordinates of each non-high-reflection point and the plane of each high-reflection region.

[0094] S250: Based on the first distance and the second distance, obtain the target non-high reflection point corresponding to each high reflection region. The target non-high reflection point is a non-high reflection point where the first distance is less than the third distance threshold and the second distance is less than the fourth distance threshold.

[0095] One approach is to compare the first distance of each of the multiple non-high reflectivity points with a third distance threshold, and to compare the second distance of each of the multiple non-high reflectivity points with a fourth distance threshold, and to obtain the target non-high reflectivity point corresponding to each high reflectivity region based on the comparison results.

[0096] Optionally, after obtaining the target non-high reflectivity point, the corresponding high reflectivity region can be expanded based on the target non-high reflectivity point, so that the expanded high reflectivity region can contain multiple data points to be processed and the target non-high reflectivity point.

[0097] In this embodiment, since some areas of the target object's surface may lack highly reflective coating, the corresponding representation in the original point cloud can be either the target area or non-highly reflective points. Therefore, expanding the highly reflective area based on the non-highly reflective points of the target can yield a more accurate target area.

[0098] S260: Project the target data points and target non-high-reflection points corresponding to each of the high-reflection regions onto the corresponding plane to obtain a two-dimensional feature map corresponding to each of the high-reflection regions.

[0099] One approach is to project the target data points and non-high-reflection points corresponding to each high-reflection region onto the corresponding plane using a vertical projection method, thereby obtaining a two-dimensional feature map corresponding to each high-reflection region.

[0100] Optionally, the unit normal vector of each plane can be obtained based on the fitting equation of the plane corresponding to each high-reflectivity region. Based on the three-dimensional coordinates of the target data points and non-high-reflectivity points corresponding to each high-reflectivity region, the fitting equation of the corresponding plane, and the normal vector, the projected coordinates of the target data points and non-high-reflectivity points corresponding to each high-reflectivity region can be obtained. A two-dimensional feature map corresponding to each high-reflectivity region can then be obtained based on the projected coordinates. The formula for calculating the projected coordinates can be:

[0101] XV i =x i -(VX*x i +VY*y i +VZ*z i +Vd)*VX-XV c

[0102] YV i =y i -(VX*x i +VY*y i +VZ*z i +Vd)*VY-YV c

[0103] ZV i =z i -(VX*x i +VY*y i +VZ*z i +Vd)*VZ-ZV c

[0104] The fitting equation for the plane corresponding to the high-reflectivity region can be:

[0105] VX*x i +VY*y i +VZ*z i +Vd=0

[0106] The normal vector of the plane corresponding to the high-reflection region can be V = (VX, VY, VZ), and the projected coordinates of the center point corresponding to the high-reflection region can be expressed as (XV...VZ). c YV c ZV c ), where XV c YV can be 0 c It can be 0, ZV c It can be 0, (x) i ,y i ,z i ) can represent the three-dimensional coordinates of the i-th data point in the high-reflection region.

[0107] Optionally, the coordinates of each projected coordinate in the two-dimensional feature map can be obtained based on the projected coordinates, the size of the two-dimensional feature map, and the resolution of a single point in the original point cloud, thus obtaining the two-dimensional feature map corresponding to each high-reflectivity region. The formula for calculating the coordinates in the two-dimensional feature map can be:

[0108]

[0109] The projected coordinates of the i-th data point in the high-reflection region can be represented as (XV) i YV i ZV i The projected coordinates of the center point corresponding to the high-reflection region can be expressed as (XV). c YV c ZV c ), where XV c YV can be 0 c It can be 0, ZV c It can be 0; W can represent the width of the 2D feature map, and H can represent the height of the 2D feature map; sp can represent the resolution of a single point in the original point cloud; e -2 It can represent 10 -2 .

[0110] S270: Input the multiple two-dimensional feature maps into a pre-trained deep learning model to obtain the boundary information corresponding to each of the multiple high-reflectivity regions. The boundary information includes shape category information and location information.

[0111] S280: Based on the shape category information and location information corresponding to each of the multiple high-reflectivity regions, a target region is obtained in each of the high-reflectivity regions, and the target region is the region where the target object is located in the original point cloud.

[0112] S290: Filter out the data points in the multiple high-reflectivity regions that are not in the corresponding target regions from the original point cloud to obtain the target point cloud.

[0113] This embodiment provides a point cloud processing method that, through the aforementioned approach, allows for the generation of multiple high-reflectivity regions and their corresponding two-dimensional feature maps based on multiple data points to be processed in the original point cloud. Based on these two-dimensional feature maps and a pre-trained deep learning model, shape category information and location information corresponding to each of the multiple high-reflectivity regions are obtained. Furthermore, based on this shape category information and location information, the region containing the target object in the original point cloud is determined. This allows for the filtering of data points other than the target region from the original point cloud, resulting in the target point cloud and improving the accuracy of noise removal from the target object's surface. In this embodiment, the high-reflectivity regions are fitted to a plane using the least squares method, and the multiple data points to be processed corresponding to the high-reflectivity regions and the target non-high-reflectivity points are projected onto corresponding planes to obtain the projected coordinates of these projections. Two-dimensional feature maps are then generated based on these projection coordinates, transforming sparse three-dimensional point cloud coordinates into dense two-dimensional coordinates. This solves the problem of difficult feature representation in 3D sparse point clouds, achieving the representation of high-reflectivity features of 3D point clouds in 2D space.

[0114] Please see Figure 10 This application provides a point cloud processing device 600, the device 600 comprising:

[0115] The data acquisition unit 610 is used to acquire an original point cloud, which includes multiple data points, each of which has a corresponding reflectance; and to acquire multiple data points to be processed, which are data points whose reflectance is greater than a reflectance threshold.

[0116] The high-reflectivity region acquisition unit 620 is used to obtain multiple high-reflectivity regions and multiple two-dimensional feature maps based on the multiple data points to be processed. The multiple high-reflectivity regions and the multiple two-dimensional feature maps are... Figure 1 In a one-to-one correspondence, each of the high-reflectivity regions is a region consisting of a set of multiple data points to be processed located in a specified circle, and each of the two-dimensional feature maps is obtained by projecting the corresponding high-reflectivity region onto a two-dimensional plane.

[0117] The target region acquisition unit 630 is used to input the multiple two-dimensional feature maps into a pre-trained deep learning model to obtain the boundary information corresponding to each of the multiple high-reflectivity regions. The boundary information includes shape category information and position information. Based on the shape category information and position information corresponding to each of the multiple high-reflectivity regions, the target region in each of the high-reflectivity regions is obtained. The target region is the region where the target object is located in the original point cloud.

[0118] The target point cloud acquisition unit 640 is used to filter out data points in the multiple high-contrast regions that are not in the corresponding target regions from the original point cloud to obtain the target point cloud.

[0119] In one approach, the high-reflectivity region acquisition unit 620 is specifically used to cluster the plurality of data points to be processed to obtain the plurality of high-reflectivity regions, wherein the center of a designated circle in each of the high-reflectivity regions is the center point of the corresponding high-reflectivity region; based on the center points corresponding to each of the plurality of high-reflectivity regions, plane fitting is performed on the plurality of high-reflectivity regions to generate two-dimensional feature maps corresponding to each of the plurality of high-reflectivity regions.

[0120] Optionally, the high-reflectivity region acquisition unit 620 is specifically used to: select one data point to be processed from multiple data points to be processed in the current processing process as an initial center point; obtain an initial set based on the data points to be processed located within a specified circle centered on the initial center point, wherein the radius of the specified circle is the initial radius; cluster all data points to be processed in the initial set to obtain high-reflectivity regions in the current processing process; delete all data points to be processed corresponding to the high-reflectivity regions in the current processing process from the multiple data points to obtain the remaining data points to be processed in the current processing process, and use the remaining data points to be processed in the current processing process as multiple data points to be processed in the next processing process, until there are no remaining data points to be processed, thereby obtaining multiple high-reflectivity regions.

[0121] Optionally, the high-reflectivity region acquisition unit 620 is specifically used to perform coarse clustering based on the initial center point and initial radius corresponding to the initial set to obtain a reference center point; to obtain an update radius based on a preset step size and the initial radius; to obtain an update set based on the data points to be processed located within a specified circle centered on the reference center point, and to use the reference center point as the initial center point of the update set and the update radius as the initial radius of the update set, wherein the radius of the specified circle is the update radius; and to perform fine clustering based on the initial center point and initial radius corresponding to the update set to obtain the high-reflectivity region in the current processing.

[0122] Optionally, the original point cloud further includes multiple non-high-reflectivity points. The high-reflectivity region acquisition unit 630 is specifically used to fit all the data points to be processed in each high-reflectivity region based on the least squares method to obtain the plane and target data points corresponding to each high-reflectivity region; to acquire a first distance and a second distance corresponding to each high-reflectivity region, wherein the first distance is the distance between each of the multiple non-high-reflectivity points and the plane corresponding to each high-reflectivity region, and the second distance is the minimum distance between each of the multiple non-high-reflectivity points and all the data points to be processed in each high-reflectivity region; based on the first distance and the second distance, to obtain the target non-high-reflectivity point corresponding to each high-reflectivity region, wherein the target non-high-reflectivity point is a non-high-reflectivity point whose first distance is less than a third distance threshold and whose second distance is less than a fourth distance threshold; and to project the target data points and target non-high-reflectivity points corresponding to each high-reflectivity region onto the corresponding plane to obtain a two-dimensional feature map corresponding to each high-reflectivity region.

[0123] Optionally, the high-reflection region acquisition unit 630 is specifically used to, during the processing of a high-reflection region, fit all the data points to be processed in the high-reflection region based on the least squares method to obtain the reference plane corresponding to the high-reflection region; if the angle between the reference plane and the specified plane is greater than the angle threshold, obtain multiple third distances, the multiple third distances being the distances between each of the data points to be processed and the reference plane; filter out interference data points from all the data points to be processed to obtain filtered data points, and continue to fit the filtered data points based on the least squares method until the angle between the reference plane corresponding to the high-reflection region and the specified plane is less than or equal to the angle threshold, and take the reference plane with the angle between it and the specified plane less than or equal to the angle threshold as the plane corresponding to the high-reflection region, the interference data points being the data points ranked first in the specified position after sorting all the data points to be processed in descending order of their respective third distances.

[0124] In one approach, the target region acquisition unit 630 is specifically used to obtain a reference high-reflection region corresponding to each of the plurality of high-reflection regions based on the shape category information and position information corresponding to each of the high-reflection regions, wherein the reference high-reflection region is the region where the target object is located in the corresponding two-dimensional feature map; and to back-project the reference high-reflection region corresponding to each of the high-reflection regions onto the original point cloud to obtain the target region in each of the high-reflection regions.

[0125] Optionally, the shape category information includes the predicted probabilities of various shapes, and the position information includes the coordinates of a pointing pixel and the indication range of the pointing pixel coordinates. The target region acquisition unit 630 is specifically used to obtain the shape of each reference high-reflectivity region based on the predicted probabilities of various shapes corresponding to each high-reflectivity region; to obtain the range of each reference high-reflectivity region based on the coordinates of a pointing pixel and the indication range of the pointing pixel coordinates corresponding to each high-reflectivity region; and to obtain the reference high-reflectivity region corresponding to each high-reflectivity region based on the shape and range of each reference high-reflectivity region.

[0126] The following will combine Figure 11 This application describes an electronic device.

[0127] Please see Figure 11 Based on the point cloud processing method and apparatus described above, this application embodiment also provides another electronic device 100 capable of executing the aforementioned point cloud processing method. The electronic device 100 includes one or more (only one shown in the figure) processors 102 and a memory 104 coupled to each other. The memory 104 stores programs capable of executing the contents of the aforementioned embodiments, and the processors 102 can execute the programs stored in the memory 104.

[0128] The processor 102 may include one or more processing cores. The processor 102 connects to various parts within the electronic device 100 using various interfaces and lines, and performs various functions and processes data of the electronic device 100 by running or executing instructions, programs, code sets, or instruction sets stored in the memory 104, and by calling data stored in the memory 104. Optionally, the processor 102 may be implemented using at least one of the following hardware forms: a Neural Network Processing Unit (NPU), a Digital Signal Processing Unit (DSP), a Field-Programmable Gate Array (FPGA), or a Programmable Logic Array (PLA). The processor 102 may integrate one or more of the following: a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), a Neural Network Processing Unit (NPU), and a modem. Specifically, the CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; the NPU handles multimedia data such as video and images; and the modem handles wireless communication. It is understandable that the aforementioned modem may not be integrated into the processor 102, but may be implemented using a separate communication chip.

[0129] The memory 104 may include random access memory (RAM) or read-only memory (ROM). The memory 104 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 104 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), and instructions for implementing the various method embodiments described below. The data storage area may also store data created by the electronic device 100 during use (such as phonebook data, audio and video data, chat log data, etc.).

[0130] This application provides a computer-readable storage medium 800. The computer-readable storage medium 800 stores program code that can be called by a processor to execute the methods described in the above method embodiments.

[0131] The computer-readable storage medium 800 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Optionally, the computer-readable storage medium 800 includes a non-transitory computer-readable storage medium. The computer-readable storage medium 800 has storage space for program code 810 that performs any of the method steps described above. This program code can be read from or written to one or more computer program products. The program code 810 may, for example, be compressed in a suitable form.

[0132] In summary, this application provides a point cloud processing method, apparatus, and electronic device that acquires multiple data points to be processed from an original point cloud, wherein the data points to be processed are data points with reflectivity greater than a reflectivity threshold; clusters the multiple data points to be processed to obtain multiple high-reflectivity regions, each of which is a region composed of a set of multiple data points to be processed located within a specified circle, the center of which is the center point of the high-reflectivity region; performs planar fitting on the multiple high-reflectivity regions based on their respective center points to generate two-dimensional feature maps corresponding to each of the multiple high-reflectivity regions; inputs the two-dimensional feature maps corresponding to each of the multiple high-reflectivity regions into a pre-trained deep learning model to obtain target regions corresponding to each of the multiple high-reflectivity regions, wherein the target regions are the regions where the target objects are located in the original point cloud; and filters out data points in the multiple high-reflectivity regions that are not within their corresponding target regions to obtain a target point cloud.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A point cloud processing method, characterized in that, The method includes: Acquire multiple data points to be processed from the original point cloud, wherein the data points to be processed are data points whose reflectance is greater than the reflectance threshold; Based on the multiple data points to be processed, multiple high-reflection regions and multiple two-dimensional feature maps are obtained. The multiple high-reflection regions correspond one-to-one with the multiple two-dimensional feature maps. Each high-reflection region is a region composed of a set of multiple data points to be processed in a specified circle. Each two-dimensional feature map is obtained by projecting the corresponding high-reflection region onto a two-dimensional plane. The multiple two-dimensional feature maps are input into a pre-trained deep learning model to obtain the boundary information corresponding to each of the multiple high-reflectivity regions. The boundary information includes shape category information and location information. Based on the shape category information and location information corresponding to each of the multiple high-reflectivity regions, a target region is obtained in each of the high-reflectivity regions, and the target region is the region where the target object is located in the original point cloud; The target point cloud is obtained by filtering out data points in the multiple high-reflectivity regions that are not in the corresponding target regions from the original point cloud.

2. The method according to claim 1, characterized in that, Based on the shape category information and location information corresponding to each of the multiple high-reflectivity regions, a target region is obtained in each of the high-reflectivity regions, including: Based on the shape category information and location information corresponding to each of the multiple high-reflectivity regions, a reference high-reflectivity region is obtained for each of the high-reflectivity regions, and the reference high-reflectivity region is the region where the target object is located in the corresponding two-dimensional feature map; The reference high-reflection region corresponding to each of the high-reflection regions is back-projected onto the original point cloud to obtain the target region in each of the high-reflection regions.

3. The method according to claim 2, characterized in that, The shape category information includes the predicted probabilities of various shapes, and the position information includes the coordinates of a pointing pixel and the indication range of the pointing pixel coordinates. The step of obtaining a reference high-reflection region corresponding to each of the multiple high-reflection regions based on the shape category information and position information of each region includes: Based on the predicted probabilities of various shapes corresponding to each of the high-reflectivity regions, the shape of each of the reference high-reflectivity regions is obtained; Based on the coordinates of the pointing pixel point corresponding to each of the high-reflectivity regions and the indication range of the pointing pixel point coordinates, the range of each of the reference high-reflectivity regions is obtained; Based on the shape and extent of each of the reference high-reflectivity regions, a reference high-reflectivity region corresponding to each of the high-reflectivity regions is obtained.

4. The method according to claim 1, characterized in that, The process of obtaining multiple high-reflectivity regions and multiple two-dimensional feature maps based on the multiple data points to be processed includes: Cluster the multiple data points to be processed to obtain multiple high-reflectivity regions, and the center of the designated circle of each high-reflectivity region is the center point of the corresponding high-reflectivity region; Based on the center points corresponding to each of the multiple high-reflectivity regions, planar fitting is performed on each of the multiple high-reflectivity regions to generate a two-dimensional feature map corresponding to each of the multiple high-reflectivity regions.

5. The method according to claim 4, characterized in that, The clustering of the multiple data points to be processed yields multiple high-reflectivity regions, including: In the current processing, one data point to be processed is selected from multiple data points to be processed in the current processing as the initial center point; An initial set is obtained based on the data points to be processed located within a specified circle centered at the initial center point, where the radius of the specified circle is the initial radius; Cluster all data points to be processed in the initial set to obtain the high-reflection regions in the current processing; All data points corresponding to the high-reflection regions in the current processing are deleted from the plurality of data points to be processed, so as to obtain the remaining data points to be processed in the current processing. The remaining data points to be processed in the current processing are used as multiple data points to be processed in the next processing, until there are no remaining data points to be processed, so as to obtain multiple high-reflection regions.

6. The method according to claim 5, characterized in that, The step of clustering all data points to be processed in the initial set to obtain high-intensity regions in the current processing includes: Based on the initial center point and initial radius corresponding to the initial set, coarse clustering is performed to obtain the reference center point; The updated radius is obtained based on the preset step size and the initial radius; An update set is obtained based on the data points to be processed located within a specified circle centered at the reference center point. The reference center point is used as the initial center point of the update set, and the update radius is used as the initial radius of the update set. The radius of the specified circle is the update radius. Based on the initial center point and initial radius corresponding to the updated set, fine clustering is performed to obtain the high-reflection region in the current processing.

7. The method according to claim 4, characterized in that, The original point cloud also includes multiple non-high-reflectivity points. The step of performing planar fitting on the multiple high-reflectivity regions based on their respective center points to generate two-dimensional feature maps for each of the multiple high-reflectivity regions includes: Based on the least squares method, all data points to be processed in each of the high-reflection regions are fitted to obtain the plane and target data points corresponding to each of the high-reflection regions; Obtain a first distance and a second distance corresponding to each of the high-reflection regions. The first distance is the distance between each of the plurality of non-high-reflection points and the plane corresponding to each of the high-reflection regions. The second distance is the minimum distance between each of the plurality of non-high-reflection points and all data points to be processed in each of the high-reflection regions. Based on the first distance and the second distance, the target non-high reflectivity point corresponding to each high reflectivity region is obtained. The target non-high reflectivity point is a non-high reflectivity point where the first distance is less than the third distance threshold and the second distance is less than the fourth distance threshold. Project the target data points and target non-high-reflection points corresponding to each high-reflection region onto the corresponding plane to obtain a two-dimensional feature map corresponding to each high-reflection region.

8. The method according to claim 7, characterized in that, The step of fitting all the data points to be processed in each of the high-reflectivity regions using the least squares method to obtain the plane and target data points corresponding to each high-reflectivity region includes: In the process of processing a high-reflectivity region, all the data points to be processed in the high-reflectivity region are fitted based on the least squares method to obtain the reference plane corresponding to the high-reflectivity region. If the angle between the reference plane and the specified plane is greater than the angle threshold, multiple third distances are obtained, where the multiple third distances are the distances between each of the data points to be processed and the reference plane; Interference data points are filtered out from all the data points to be processed to obtain filtered data points. The filtered data points are then fitted using the least squares method until the angle between the reference plane corresponding to the high-reflectivity region and the specified plane is less than or equal to the angle threshold. The reference plane with an angle less than or equal to the angle threshold is taken as the plane corresponding to the high-reflectivity region. The interference data points are the data points ranked first in the specified positions after all the data points to be processed are sorted in descending order of their respective third distances.

9. A point cloud processing device, characterized in that, The device includes: A data acquisition unit is used to acquire a raw point cloud, which includes multiple data points, each of which has a corresponding reflectance; and to acquire multiple data points to be processed, which are data points whose reflectance is greater than a reflectance threshold. The high-reflection region acquisition unit is used to obtain multiple high-reflection regions and multiple two-dimensional feature maps based on the multiple data points to be processed. The multiple high-reflection regions correspond one-to-one with the multiple two-dimensional feature maps. Each high-reflection region is a region composed of a set of multiple data points to be processed in a specified circle. Each two-dimensional feature map is obtained by projecting the corresponding high-reflection region onto a two-dimensional plane. The target region acquisition unit is used to input the multiple two-dimensional feature maps into a pre-trained deep learning model to obtain the boundary information corresponding to each of the multiple high-reflectivity regions. The boundary information includes shape category information and position information. Based on the shape category information and position information corresponding to each of the multiple high-reflectivity regions, the target region in each of the high-reflectivity regions is obtained. The target region is the region where the target object is located in the original point cloud. The target point cloud acquisition unit is used to filter out data points in the multiple high-reflectivity regions that are not in the corresponding target regions from the original point cloud to obtain the target point cloud.

10. An electronic device, characterized in that, Includes one or more processors and memory; One or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs being configured to perform the method of any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code, wherein the method described in any one of claims 1-8 is executed when the program code is run.

Images