Laser Point Cloud Denoising Method and System

Abstract

An embodiment of the present invention discloses a method and system for laser point cloud denoising. The method includes: preprocessing frame point clouds to obtain initial point clouds; deleting points with close distances and low reflectivity in the initial point clouds and screening out a first set of points to be processed; searching for surrounding points for each point in the first set of points to be processed in the processed point clouds, determining target noise points, attached points, and target surrounding points corresponding to the target noise points, adding the target surrounding points to the first set of points to be processed to obtain a second set of points to be processed, and updating the attached points; clustering the second set of points to be processed and determining noise point clusters, adding the points therein to the initial noise point set to obtain the latest noise point set; performing noise marking on the intrusion point clouds according to the latest noise point set, then clustering and judging effective clusters and determining noise points, deleting the corresponding noise points to be removed from the effective clusters to obtain target clusters; updating the current background point cloud according to the processed point cloud and the current background point cloud. The accuracy of denoising and the target recognition accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the field of lidar, and in particular to a method and system for denoising lidar point clouds. Background Art

[0002] Based on the characteristic of being independent of ambient light, three-dimensional lidar point clouds are often used to replace image detectors for target detection in low-illumination environments. High-precision lidar can be more applied to perimeter intrusion detection with a higher security level. However, the detection principle of lidar also determines that its detection range will weaken in weather such as rain, snow, and fog, and noise points will be generated, resulting in false alarms. The industry's application of lidar often focuses on autonomous driving. The solutions in this scenario only consider non-severe weather such as light rain, snow, and fog, and are not applicable to the working requirements in security scenarios such as heavy rain and thick fog. At the same time, in some other security scenarios, aerosol foreign matters such as dust and water vapor will also cause more complex false alarm situations. In some existing other algorithms, most of them only perform simple screening based on distribution or reflectivity characteristics. The screening process depends on environmental characteristics and has poor robustness; or only the removal of noise points is considered, and the possibility that the incident angle or object material may also generate similar feature point clouds is ignored.

[0003] For the above problems in the prior art, there is currently no effective solution. Summary of the Invention

[0004] To solve the above problems, the present invention provides a method and system for denoising lidar point clouds. By analyzing the reflectivity, density, distribution, etc., noise points are determined round by round. At the same time, the noise points attached to the object edges are restored, and the background is updated in real time to solve the problems of poor denoising effect and low accuracy in the prior art.

[0005] To achieve the above object, the present invention provides a method for denoising laser point clouds, including: obtaining the frame point cloud of the current frame, and obtaining the initial point cloud after preprocessing the frame point cloud; moving the points with a reflectivity lower than the first preset reflectivity threshold and a distance from the origin less than the preset distance threshold from the initial point cloud to the initial noise point set to obtain the processed point cloud; taking the points with a reflectivity lower than the second preset reflectivity threshold and a distance from the origin less than the preset distance threshold in the processed point cloud as the first set of points to be processed; wherein, the second preset reflectivity threshold is greater than the first preset reflectivity threshold; traversing the first set of points to be processed, performing neighborhood search on each point to be processed in the processed point cloud according to the preset radius to obtain surrounding points, and judging whether the point to be processed is a target noise point according to the distribution characteristics and reflectivity of the surrounding points. If so, determining the target surrounding points corresponding to the target noise point according to the reflectivity, and adding the target surrounding points to the first set of points to be processed; if not, taking the point to be processed as an attachment point and moving it from the first set of points to be processed to the attachment point set, and updating the position and reflectivity of the attachment point in the processed point cloud according to the corresponding surrounding points; obtaining the second set of points to be processed after traversing; clustering the second set of points to be processed to obtain the first clustering result; judging the noise clustering according to the clustering characteristics of the first clustering result, adding the points in the noise clustering to the initial noise point set to obtain the latest noise point set; performing noise marking on the intrusion point cloud according to the latest noise point set; clustering the intrusion point cloud to obtain the second clustering result, judging the effective clustering and the noise points to be removed according to the clustering characteristics of the second clustering result, and deleting the corresponding noise points to be removed from the effective clustering to obtain the target clustering; canceling the marking of the remaining noise points in the target clustering, and deleting the remaining noise points from the latest noise point set; updating the current background point cloud according to the processed point cloud and the current background point cloud; wherein, the intrusion point cloud is obtained by differentiating the processed point cloud and the current background point cloud in advance.

[0006] Further optionally, the judging whether the point to be processed is a target noise point according to the distribution characteristics and reflectivity of the surrounding points, and if so, determining the target surrounding points corresponding to the target noise point according to the reflectivity, includes: retrieving the surrounding points with a reflectivity higher than or equal to the second preset reflectivity threshold, denoted as the first surrounding points, counting the number and positions of the first surrounding points. When the number of the first surrounding points is 0, taking the corresponding point to be processed as the target noise point; retrieving the surrounding points with a reflectivity lower than the second preset reflectivity threshold, denoted as the second surrounding points, counting the number and positions of the second surrounding points, calculating the distribution density of the second surrounding points according to the number of the second surrounding points, and taking the point to be processed corresponding to the distribution density lower than the first preset density threshold as the target noise point; calculating the first reflectivity mean value and the first standard deviation of the reflectivities of all the surrounding points corresponding to each point to be processed, and taking the point to be processed corresponding to the first standard deviation lower than the first preset standard deviation threshold as the target noise point; taking the surrounding points with a reflectivity lower than the first reflectivity mean value as the target surrounding points corresponding to the target noise point.

[0007] Further optionally, updating the position and reflectivity of the attachment point in the processed point cloud according to the corresponding surrounding points includes: calculating the second reflectivity average of the first surrounding points corresponding to each attachment point, and using the second reflectivity average as the updated reflectivity of the corresponding attachment point; calculating the quantity ratio of the first surrounding points to the second surrounding points, and calculating the reflectivity ratio of the updated reflectivity to the unupdated reflectivity; calculating the first center position of the first surrounding point and the second center position of the second surrounding point; and calculating the updated position according to the quantity ratio, the reflectivity ratio, the first center position and the second center position.

[0008] Further optionally, the method of determining noise clusters based on the clustering characteristics of the first clustering result includes: calculating the length of each cluster in the first clustering result in depth, calculating the third reflectivity mean and second standard deviation of all points in the cluster; treating clusters whose length is less than a first preset length threshold, whose third reflectivity mean is less than the first preset reflectivity threshold, and whose second standard deviation is less than the second preset standard deviation threshold as noise clusters; treating clusters whose number of points is less than a preset number threshold, or whose point density is less than a second preset density threshold as noise clusters; calculating Gaussian curvature after thinning the clusters, calculating a first quantity ratio of points whose Gaussian curvature is lower than the preset curvature threshold in the cluster to which they belong, and when the first quantity ratio is greater than the first preset ratio threshold, treating the corresponding cluster as a noise cluster.

[0009] Further optionally, the determining of valid clusters and noise points to be removed based on the clustering characteristics of the second clustering result includes: counting the number of noise points and the proportion of the number of noise points in each cluster in the second clustering result; taking a cluster whose number of noise points is lower than a preset noise point number threshold, or a cluster whose proportion of the number of noise points is lower than a preset noise point number proportion threshold as a valid cluster, and taking the noise points in the valid cluster as noise points to be removed; calculating the point distance from each noise point in the cluster to the nearest non-noise point, and calculating the second number proportion of points whose point distance is greater than or equal to a second preset length threshold in the cluster to which it belongs, when the second number proportion is greater than the second preset proportion threshold, taking the corresponding cluster as an invalid cluster, otherwise it is a valid cluster; taking the points in the valid cluster whose point distance is less than the second preset length threshold as points to be updated, calculating the position and reflectivity to be updated based on the corresponding surrounding points of the points to be updated, and judging whether there are points whose distance exceeds the second preset length threshold within the preset radius of the position to be updated, if not, using the position to be updated and reflectivity to be updated to respectively update the position and reflectivity of the point to be updated; and taking the unupdated points to be updated as noise points to be removed.

[0010] Further optionally, updating the current background point cloud according to the processed point cloud and the current background point cloud includes: counting the proportion of noise points of the points in the processed point cloud within a first preset distance threshold range, and when the proportion of noise points is greater than a third preset proportion threshold, not updating the current background point cloud; otherwise, constructing a first depth map of the initial point cloud and a second depth map of the current background point cloud, and performing grid division on the first depth map and the second depth map; for each first grid in the first depth map, when there are noise points therein, adding all the points in the second grid corresponding to the first grid in the second depth map to the updated background point set, and adding all the points in the eight grids around the corresponding second grid to the updated background point set; when there are no noise points and object points in the first grid, adding all the points in the first grid to the updated background point set; when there are object points but no noise points in the first grid, adding the non-object points in the first grid and all the points in the corresponding second grid to the updated background point set to obtain the updated background point cloud.

[0011] On the other hand, the present invention also provides a laser point cloud denoising system, including: a preprocessing module, configured to obtain the frame point cloud of the current frame, and obtain the initial point cloud after preprocessing the frame point cloud; an initial denoising module, configured to move the points in the initial point cloud whose reflectivity is lower than the first preset reflectivity threshold and whose distance from the origin is less than the preset distance threshold to the initial noise point set to obtain the processed point cloud; and use the points in the processed point cloud whose reflectivity is lower than the second preset reflectivity threshold and whose distance from the origin is less than the preset distance threshold as the first set of points to be processed; wherein the second preset reflectivity threshold is greater than the first preset reflectivity threshold; a first noise point screening module, configured to traverse the first set of points to be processed, perform a neighborhood search in the processed point cloud for each point to be processed according to a preset radius to obtain surrounding points, and determine whether the point to be processed is a target noise point according to the distribution characteristics and reflectivity of the surrounding points. If so, determine the target surrounding points corresponding to the target noise point according to the reflectivity, and add the target surrounding points to the first set of points to be processed; if not, move the point to be processed as an attachment point from the first set of points to be processed to the attachment point set, and update the position and reflectivity of the attachment point in the processed point cloud according to the corresponding surrounding points; after traversal, obtain the second set of points to be processed; a second noise point screening module, configured to perform clustering on the second set of points to be processed to obtain a first clustering result; determine noise point clustering according to the clustering characteristics of the first clustering result, and add the points in the noise point clustering to the initial noise point set to obtain the latest noise point set; a noise point removal module, configured to perform noise point marking on the intrusion point cloud according to the latest noise point set; perform clustering on the intrusion point cloud to obtain a second clustering result, determine valid clustering and noise points to be removed according to the clustering characteristics of the second clustering result, delete the corresponding noise points to be removed from the valid clustering to obtain the target clustering; cancel the marking of the remaining noise points in the target clustering, and delete the remaining noise points from the latest noise point set, and perform target recognition according to the target clustering; update the current background point cloud according to the processed point cloud and the current background point cloud; wherein the intrusion point cloud is obtained by differentiating the processed point cloud from the current background point cloud in advance.

[0012] Further optionally, the first noise screening module includes: a first target noise determination sub-module, configured to retrieve surrounding points with a reflectivity higher than or equal to a second preset reflectivity threshold, denoted as first surrounding points, count the number and positions of the first surrounding points, and when the number of the first surrounding points is 0, use the corresponding point to be processed as a target noise; a second target noise determination sub-module, configured to retrieve surrounding points with a reflectivity lower than the second preset reflectivity threshold, denoted as second surrounding points, count the number and positions of the second surrounding points, calculate the distribution density of the second surrounding points according to the number of the second surrounding points, and use the point to be processed corresponding to the distribution density lower than a first preset density threshold as a target noise; a third target noise determination sub-module, configured to calculate a first reflectivity mean value and a first standard deviation of the reflectivities of all surrounding points corresponding to each point to be processed, and use the point to be processed corresponding to the first standard deviation lower than a first preset standard deviation threshold as a target noise; a target surrounding point determination sub-module, configured to use the surrounding points with a reflectivity lower than the first reflectivity mean value as target surrounding points corresponding to the target noise.

[0013] Further optionally, the second noise screening module includes: a calculation sub-module, configured to calculate the length of each cluster in the first clustering result in terms of depth, and calculate a third reflectivity mean value and a second standard deviation of all points within the cluster; a first noise cluster determination sub-module, configured to use the cluster with a length less than a first preset length threshold, a third reflectivity mean value less than a first preset reflectivity threshold, and a second standard deviation less than a second preset standard deviation threshold as a noise cluster; a second noise cluster determination sub-module, configured to use the cluster with the number of points less than a preset number threshold or the point density less than a second preset density threshold as a noise cluster; a third noise cluster determination sub-module, configured to thin out the cluster and then calculate the Gaussian curvature, calculate the first quantity proportion of the points with a Gaussian curvature lower than a preset curvature threshold within the cluster to which they belong, and when the first quantity proportion is greater than a first preset proportion threshold, use the corresponding cluster as a noise cluster.

[0014] Further optionally, the noise removal module includes: a first denoising sub-module, configured to count the number of noise points and the proportion of noise points in each cluster in the second clustering result; taking the clusters with the number of noise points lower than a preset noise point number threshold, or the clusters with the proportion of noise points lower than a preset noise point proportion threshold as valid clusters, and taking the noise points in the valid clusters as the noise points to be removed; a valid cluster determination sub-module, configured to calculate the point distance from each noise point in the cluster to the nearest non-noise point, calculate the second proportion of the number of points with the point distance greater than or equal to a second preset length threshold in the cluster to which they belong, and when the second proportion is greater than a second preset proportion threshold, taking the corresponding cluster as an invalid cluster, otherwise as a valid cluster; a second denoising sub-module, configured to take the points with the point distance less than the second preset length threshold in the valid cluster as the points to be updated, calculate the position and reflectivity to be updated according to the surrounding points corresponding to the points to be updated, and determine whether there are points with a distance exceeding the second preset length threshold within a preset radius of the position to be updated. If not, updating the position and reflectivity of the points to be updated with the position and reflectivity to be updated respectively; taking the points to be updated that are not updated as the noise points to be removed.

[0015] The above technical solution has the following beneficial effects: By analyzing information such as the distribution and reflectivity of points in the initial point cloud, noise points are gradually filtered, and possible noise points are processed one by one according to their causes to improve the accuracy of noise removal caused by extreme weather; The noise points at the edges of objects are screened out and restored to the contour of the object itself as attachments, which is convenient for subsequent target recognition and improves the accuracy of subsequent target recognition; The background is selectively updated, and the true situation of the background is restored according to the screened noise points, thereby ensuring the accuracy of the intrusion point cloud after background difference and improving the accuracy of subsequent target recognition; Traditional algorithms are used to improve the operation speed and ensure real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to these drawings.

[0017] Figure 1 is a flowchart of the laser point cloud denoising method provided by the embodiment of the present invention;

[0018] Figure 2 is a flowchart of the target noise point determination method provided by the embodiment of the present invention;

[0019] Figure 3 is a flowchart of the attachment point update method provided by the embodiment of the present invention;

[0020] Figure 4 It is a flowchart of the noise point clustering judgment method provided by an embodiment of the present invention;

[0021] Figure 5 It is a flowchart of the effective clustering and noise points to be removed judgment method provided by an embodiment of the present invention;

[0022] Figure 6 It is a flowchart of the background update method provided by an embodiment of the present invention;

[0023] Figure 7 It is a schematic structural diagram of the laser point cloud denoising system provided by an embodiment of the present invention;

[0024] Figure 8 It is a schematic structural diagram of the first noise point screening module for target noise point determination provided by an embodiment of the present invention;

[0025] Figure 9 It is a schematic structural diagram of the first noise point screening module for attachment point update provided by an embodiment of the present invention;

[0026] Figure 10 It is a schematic structural diagram of the second noise point screening module provided by an embodiment of the present invention;

[0027] Figure 11 It is a schematic structural diagram of the noise point removal module for denoising provided by an embodiment of the present invention;

[0028] Figure 12 It is a schematic structural diagram of the noise point removal module for background reconstruction provided by an embodiment of the present invention.

[0029] Reference numerals: 100 - preprocessing module; 200 - primary denoising module; 300 - first noise point screening module; 3001 - first target noise point determination sub - module; 3002 - second target noise point determination sub - module; 3003 - third target noise point determination sub - module; 3004 - target peripheral point determination sub - module; 3005 - reflectivity update sub - module; 3006 - reflectivity ratio calculation sub - module; 3007 - center position calculation sub - module; 3008 - position update sub - module; 400 - second noise point screening module; 4001 - calculation sub - module; 4002 - first noise point clustering determination sub - module; 4003 - second noise point clustering determination sub - module; 4004 - third noise point clustering determination sub - module; 500 - noise point removal module; 5001 - first denoising sub - module; 5002 - effective clustering determination sub - module; 5003 - second denoising sub - module; 5004 - update judgment sub - module; 5005 - first grid processing sub - module; 5006 - second grid processing sub - module; 5007 - third grid processing sub - module. Detailed implementation manners

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] In order to solve the problem that the existing technology has poor effect in removing noise (the noise specifically refers to the noise caused by rain, snow, fog, dust, etc. mentioned in the background technology), Figure 1 is a flow chart of a laser point cloud denoising method provided by an embodiment of the present invention. Figure 1 As shown, an embodiment of the present invention provides a laser point cloud denoising method, comprising:

[0032] S1. Obtain the frame point cloud of the current frame, and obtain the initial point cloud after preprocessing the frame point cloud.

[0033] The original point cloud data of the current frame, namely the frame point cloud, is collected by the laser radar. All points in the frame point cloud are in the same three-dimensional coordinate system with the laser radar as the origin, and each point has a three-dimensional coordinate.

[0034] The frame point cloud is an unprocessed point cloud. In order to speed up the calculation, it can be preprocessed before the noise is marked to obtain the initial point cloud. The preprocessing includes direct filtering the frame point cloud according to the security area, cutting out the effective defense area, and eliminating the data of distant or irrelevant areas; then, the ground fitting result is calculated to remove the point cloud on the ground and below the ground.

[0035] S2. Move points in the initial point cloud whose reflectivity is lower than a first preset reflectivity threshold and whose distance from the origin is less than a preset distance threshold to the initial noise point set to obtain a processed point cloud; take points in the processed point cloud whose reflectivity is lower than a second preset reflectivity threshold and whose distance from the origin is less than the preset distance threshold as the first point set to be processed; wherein the second preset reflectivity threshold is greater than the first preset reflectivity threshold.

[0036] Radar will form noise points with lower reflectivity that are obviously different from physical objects in rain, snow, fog and other environments. However, points with lower reflectivity are not just noise points such as rain, snow, fog, etc. They may also be caused by the incident angle of the laser or the material of the object. Therefore, they cannot be simply judged by the reflectivity value. For example, fog, dust, etc. may also adhere to the object, causing the point cloud of the object to be eroded or generating noise points with slightly higher reflectivity. Based on the above reasons, it is necessary to conduct a detailed analysis and judgment of the noise points.

[0037] Traverse all points in the initial point cloud and select points with reflectivity lower than the first preset reflectivity threshold R 1 And the distance to the origin (lidar) is less than the preset distance threshold D1 The discrete points are recorded as the initial noise point set, and these discrete points are removed from the initial point cloud to obtain the processed point cloud. Since severe noise conditions only occur in positions relatively close to the lidar, the low-reflectivity points close to the origin can be directly identified as noise points. Among them, the reflectivity can be directly obtained from the lidar data.

[0038] Then, mark the points with a reflectivity lower than the second preset reflectivity threshold R 2 and a distance from the origin less than the preset distance threshold D 1 These points are recorded as the to-be-processed point set, and the points in the to-be-processed point set need to be further determined whether they are noise points or attachment points. The second preset reflectivity threshold R 2 is slightly greater than the first preset reflectivity threshold R 1 to obtain all possible noise points.

[0039] The above first preset reflectivity threshold, second preset reflectivity threshold, and preset distance threshold can all be manually set according to the actual situation.

[0040] S3. Traverse the first to-be-processed point set, perform a neighborhood search for each to-be-processed point in the processed point cloud according to the preset radius to obtain surrounding points, and judge whether the to-be-processed point is a target noise point according to the distribution characteristics and reflectivity of the surrounding points. If so, determine the target surrounding points corresponding to the target noise point according to the reflectivity, and add the target surrounding points to the first to-be-processed point set; if not, move the to-be-processed point as an attachment point from the first to-be-processed point set to the attachment point set, and update the position and reflectivity of the attachment point in the processed point cloud according to the corresponding surrounding points; after traversing, obtain the second to-be-processed point set.

[0041] For each to-be-processed point in the first to-be-processed point set, use the preset radius D 2 to retrieve and view the distribution characteristics and reflectivity of its surrounding points in the processed point cloud, and perform a preliminary screening of noise points and attachment points according to the distribution characteristics and reflectivity of the surrounding points. For the obtained target noise points, if a to-be-processed point is determined to be a target noise point, then select the surrounding points suspected of being noise points among the target surrounding points corresponding to the target noise point according to the reflectivity, that is, the target surrounding points, and add the target surrounding points to the first to-be-processed point set; otherwise, this point may be an attachment point, add this point to the attachment point set and remove it from the first to-be-processed point set. Traverse all the points in the first to-be-processed point set, and judge the situation of each to-be-processed point using the above method to obtain the second to-be-processed point set and the attachment point set.

[0042] For each point in the attachment point set, it is necessary to restore it to the shape of the object it belongs to, update its position and reflectivity through the distribution characteristics and reflectivity of its surrounding points, and update the updated position result and reflectivity result to the processed point cloud.

[0043] Among them, the preset radius is set manually according to the actual clearance.

[0044] S4. Cluster the second point set to be processed to obtain the first clustering result; judge the noise clusters according to the clustering characteristics of the first clustering result, and add the points in the noise clusters to the initial noise point set to obtain the latest noise point set.

[0045] First, construct a depth map (range Image) of the second point set to be processed. If the scanning method of the hardware is petal-shaped scanning, the angles between the points in the scanning result are fixed, so the depth map can be used to ensure that the interval between the scanning points is not uneven due to the change in the distance between the object and the sensor. Of course, if it is a method such as row-by-row scanning, the projection map of the YZ plane can also be used.

[0046] Cluster based on the depth map to obtain the first clustering result, and calculate the clustering characteristics of each cluster in the first clustering result, such as the distribution of points and reflectivity in the cluster. According to the clustering characteristics, it can be judged whether each cluster is a noise cluster. If a certain cluster is a noise cluster, the points in the cluster are added to the initial noise point set to obtain the latest noise point set. Through the above method, the noise can be further screened to improve the accuracy of noise marking.

[0047] S5. Mark the noise points in the intrusion point cloud according to the latest noise point set; cluster the intrusion point cloud to obtain the second clustering result, judge the valid clusters and the noise points to be removed according to the clustering characteristics of the second clustering result, delete the corresponding noise points to be removed from the valid clusters to obtain the target clusters; cancel the marking of the remaining noise points in the target clusters, and delete the remaining noise points from the latest noise point set, and perform target recognition according to the target clusters; update the current background point cloud according to the processed point cloud and the current background point cloud; among them, the intrusion point cloud is obtained in advance by differentiating the processed point cloud and the current background point cloud.

[0048] The intrusion point cloud can be obtained after obtaining the processed point cloud. Specifically, the possible intrusion point cloud is segmented by means of background difference and pre-order difference between the processed point cloud and the current background point cloud. Mark the noise points in the intrusion point cloud according to the positions of the points in the latest noise point set.

[0049] Cluster the intrusion point cloud to obtain the second clustering result. The clustering can use the Euclidean clustering method based on distance. During the clustering process, it should be noted that the points marked as noise can be added to a certain cluster but not used as the starting point for the next adjacent retrieval to be pushed onto the stack.

[0050] Traverse each cluster in the second clustering result, determine whether it is a valid cluster, and identify the noise points to be removed therein. If it is a valid cluster, delete the noise points to be removed therein to obtain a target cluster without noise points. Cancel the marking of the remaining noise points within the target cluster and delete the remaining noise points from the latest noise point set. For the obtained target cluster, subsequent target recognition processes can be performed to obtain recognition results. The target recognition process can be completed using a recognition model, and the target cluster can be updated after recognition.

[0051] Furthermore, background point clouds are required in the differential process, and the accuracy of the background point clouds will also affect the final recognition accuracy. For example, relatively thick fog or dust sometimes causes certain occlusion of the background, resulting in the background being revealed again after the fog or dust moves and being misrecognized as an intrusion object. Therefore, it is necessary to construct a real-time updated background to assist in the differential. Combine the processed point cloud with the current background point cloud according to the noise point distribution in the processed point cloud to obtain an updated background point cloud for use in the differential process of the next frame.

[0052] As an alternative implementation Figure 2 is the flowchart of the target noise point determination method provided by the embodiment of the present invention. As shown in Figure 2 it is determined whether the point to be processed is a target noise point according to the distribution characteristics and reflectivity of the surrounding points. If so, determine the target surrounding points corresponding to the target noise point according to the reflectivity, including:

[0053] S301. Retrieve the surrounding points with a reflectivity higher than or equal to the second preset reflectivity threshold, denoted as the first surrounding points. Count the number and positions of the first surrounding points. When the number of the first surrounding points is 0, regard the corresponding point to be processed as a target noise point.

[0054] For each point to be processed, count the surrounding points with a reflectivity greater than or equal to the second preset reflectivity threshold R 2 among its corresponding surrounding points, regard them as the first surrounding points, and count the number and positions of the first surrounding points. If the number of the first surrounding points is 0, or the number of the first surrounding points is lower than a certain threshold, then consider this point to be definitely a noise point, denoted as a target noise point.

[0055] S302. Retrieve the surrounding points with a reflectivity lower than the second preset reflectivity threshold, denoted as the second surrounding points. Count the number and positions of the second surrounding points. Calculate the distribution density of the second surrounding points according to the number of the second surrounding points, and regard the point to be processed corresponding to the distribution density lower than the first preset density threshold as a target noise point.

[0056] For each point to be processed, count the surrounding points with a reflectivity less than the second preset reflectivity threshold R 2The peripheral points are used as the second peripheral points, and the number and positions of the second peripheral points are counted. According to the number of the second peripheral points, the distribution density of the low-reflectivity points around the point to be processed can be calculated. If the corresponding distribution density of the point to be processed is lower than the first preset density threshold T 1 , it is considered that this point must be a noise point and is recorded as the target noise point.

[0057] S303. Calculate the first reflectivity mean value and the first standard deviation of the reflectivities of all the peripheral points corresponding to each point to be processed, and use the point to be processed corresponding to the first standard deviation lower than the first preset standard deviation threshold as the target noise point.

[0058] For each point to be processed, calculate the first reflectivity mean value of its peripheral points and the standard deviation of the reflectivities. If the standard deviation is lower than the first preset standard deviation threshold, it is considered that this point to be processed must be a noise point and is recorded as the target noise point.

[0059] As an optional implementation manner, the first preset standard deviation threshold is usually 0.5×R 2 .

[0060] For a point to be processed, as long as one of the above three judgment conditions is met, it is considered to be the target noise point.

[0061] S304. Use the peripheral points with reflectivities lower than the first reflectivity mean value as the target peripheral points corresponding to the target noise points.

[0062] If a certain point to be processed is the target noise point, use the peripheral points with reflectivities less than the first reflectivity mean value among its peripheral points as the target peripheral points and add them to the first set of points to be processed.

[0063] As an optional implementation manner, Figure 3 is the flowchart of the attachment point update method provided by the embodiment of the present invention. As Figure 3 shown, update the position and reflectivity of the attachment point in the point cloud to be processed according to the corresponding peripheral points, including:

[0064] S305. Calculate the second reflectivity mean value of the first peripheral points corresponding to each attachment point, and use the second reflectivity mean value as the updated reflectivity of the corresponding attachment point.

[0065] For each attachment point in the attachment point set, calculate the reflectivity mean value of its corresponding first peripheral points, that is, the second reflectivity mean value, as the new reflectivity of this attachment point.

[0066] S306. Calculate the ratio of the number of the first peripheral points to the number of the second peripheral points, and calculate the ratio of the updated reflectivity to the non-updated reflectivity.

[0067] For each attachment point, calculate the ratio r of the number of the first peripheral points to the number of the second peripheral points 附着1, and calculate the reflectivity ratio r of the updated reflectivity to the original reflectivity 附着2 .

[0068] S307. Calculate the first central position of the first peripheral point and the second central position of the second peripheral point;

[0069] Calculate the first central position c of the first peripheral point respectively 附着1 and the second central position c of the second peripheral point 附着1 . Since the positions of each first peripheral point and each second peripheral point have been counted before, the corresponding central positions can be directly calculated through the position coordinates.

[0070] S308. Calculate the updated position according to the quantity ratio, the reflectivity ratio, the first central position and the second central position.

[0071] As an optional implementation manner, calculate the updated position c according to the quantity ratio, the reflectivity ratio, the first central position and the second central position 附着 by the following formula:

[0072] c 附着 = c 附着1 +(c 附着2 - c 附着1 ) × (r 附着1 × w 1 + r 附着2 × w 2 ). Where w 1 and w 2 are specified weights.

[0073] As an optional implementation manner, Figure 4 is the flowchart of the noise point clustering judgment method provided by the embodiment of the present invention. As shown in Figure 4 , judge the noise point clustering according to the clustering characteristics of the first clustering result, including:

[0074] S401. Calculate the length of each clustering in the first clustering result in terms of depth, and calculate the third reflectivity mean value and the second standard deviation of all points within the clustering.

[0075] For each clustering in the first clustering result, calculate its length in terms of depth, calculate the third reflectivity mean value and the second standard deviation of the points inside it, count the number of points inside the clustering, and calculate the point density inside the clustering based on the quantity.

[0076] S402. Take the clustering with a length less than the first preset length threshold, a third reflectivity mean value less than the first preset reflectivity threshold, and a second standard deviation less than the second preset standard deviation threshold as the noise point clustering;

[0077] For each cluster, if its corresponding length is less than the first preset length threshold L 1 , and the third reflectivity mean is less than the first preset reflectivity threshold R 1 , and the second standard deviation is less than the second preset standard deviation threshold (usually 0.5×R 1 ), then the cluster is considered to be a noise cluster, that is, the points in it must be noise points, and all the points in the cluster are added to the initial noise point set.

[0078] S403, treating clusters whose point quantity is less than a preset quantity threshold, or clusters whose point density is less than a second preset density threshold, as noise clusters;

[0079] For each cluster, if the number of points in the cluster is less than the preset number threshold, or the point density is less than the second preset density threshold, it is considered that the points in the cluster may be noise points but are relatively discrete and may not affect subsequent detection. The cluster is treated as a noise cluster and the points in the cluster are added to the initial noise point set.

[0080] S404, calculating Gaussian curvature after thinning the clusters, calculating a first quantity ratio of points whose Gaussian curvature is lower than a preset curvature threshold in the corresponding cluster, and when the first quantity ratio is greater than a first preset ratio threshold, treating the corresponding cluster as a noise point cluster.

[0081] The remaining clusters after the above two judgment methods may be noise clusters and need further judgment. First, the points in the cluster are thinned out, the Gaussian curvature of the thinned points is calculated, the number of points whose Gaussian curvature is lower than the preset curvature threshold is counted, and the first number ratio of these points to the number of all points in the cluster to which they belong is calculated. If the ratio is too high (greater than the first preset ratio threshold), the cluster is considered to be a noise cluster and the points in it are added to the initial noise point set; otherwise, the cluster is not a noise cluster.

[0082] As an optional implementation, Figure 5 is a flow chart of an effective clustering and noise point determination method to be removed provided by an embodiment of the present invention. Figure 5 As shown, judging effective clusters and noise points to be removed according to the clustering features of the second clustering result includes:

[0083] S501. Count the number of noise points and the proportion of noise points in each cluster in the second clustering result; take the clusters whose number of noise points is lower than a preset noise point number threshold, or the clusters whose proportion of noise points is lower than a preset noise point number proportion threshold as valid clusters, and take the noise points in the valid clusters as noise points to be removed.

[0084] For each cluster in the second clustering result, if there is no noise point in the cluster (the number of noise points is lower than the preset noise point number threshold) or the proportion of noise points is small (the proportion of noise points is lower than the preset noise point proportion threshold), the cluster is considered a valid cluster, and all the marked noise points in the cluster are the noise points to be removed. Remove these noise points to be removed from the valid cluster.

[0085] S502. Calculate the point distance from each noise point in the cluster to the nearest non-noise point, calculate the second quantity proportion of the points with the point distance greater than or equal to the second preset length threshold in the cluster to which they belong. When the second quantity proportion is greater than the second preset proportion threshold, the corresponding cluster is regarded as an invalid cluster, otherwise it is a valid cluster.

[0086] For the remaining clusters, calculate the point distance from each noise point to the nearest non-noise point, count the number of points with the point distance greater than or equal to the second preset length threshold D 3 and calculate its second quantity proportion to the total number of points in the cluster. If the second quantity proportion is relatively high (the second quantity proportion is greater than the second preset proportion threshold), then the cluster is considered invalid, otherwise it is valid.

[0087] S503. Take the points with the point distance less than the second preset length threshold in the valid cluster as the points to be updated. Calculate the position and reflectivity to be updated according to the surrounding points corresponding to the points to be updated. Determine whether there are points with a distance exceeding the second preset length threshold within the preset radius of the position to be updated. If not, update the position and reflectivity of the points to be updated with the position and reflectivity to be updated respectively; take the unupdated points to be updated as the noise points to be removed.

[0088] For the valid clusters determined to be valid, take the points with the point distance less than the second preset length threshold D 3 as the attachment points (points to be updated). The attachment points need to be updated. Calculate the position to be updated and the reflectivity to be updated according to the distribution and reflectivity of the surrounding points of the points to be updated. The specific update method can be the same as the update method of the above attachment points.

[0089] Determine whether there are points with a distance greater than the second preset length threshold D 2 within the preset radius D 3 near the updated position. If not, update with the position to be updated and the reflectivity to be updated, otherwise do not update. Take the unupdated points to be updated as the noise points to be removed, and remove all the noise points to be removed within the valid cluster. Cancel the noise point marking for the remaining noise points (updated noise points) within the valid cluster and delete these points from the latest noise point set.

[0090] As an alternative implementation manner, Figure 6 is the flowchart of the background update method provided by the embodiment of the present invention, as Figure 6As shown, updating the current background point cloud according to the processed point cloud and the current background point cloud includes:

[0091] S504. Statistically calculate the proportion of noise points of the points in the processed point cloud within the range of the first preset distance threshold. When the proportion of noise points is greater than the third preset proportion threshold, do not update the current background point cloud; otherwise, construct the first depth map of the initial point cloud and the second depth map of the current background point cloud, and perform grid division on the first depth map and the second depth map.

[0092] When attempting to construct the background each time, statistically calculate the proportion of noise points of the processed point cloud within the range of the first preset distance threshold. If the proportion is too large (the proportion of noise points is greater than the third preset proportion threshold), give up updating the current background point cloud; otherwise, perform the update.

[0093] When updating, first construct the first depth map Image of the processed point cloud 1 , and at the same time construct the second depth map Image of the current background point cloud 0 . Perform grid division on the first depth map and the second depth map according to the same rules.

[0094] S505. For each first grid in the first depth map, when there are noise points in it, add all the points in the second grid corresponding to the first grid in the second depth map to the updated background point set, and add all the points in the eight grids around the corresponding second grid to the updated background point set.

[0095] For each first grid in the first depth map, when there are noise points in a certain first grid, add all the points in the second grid corresponding to the first grid in the second depth map Image 0 to the updated background point set P back , and consider that the eight grids near the second grid are all added to the background point set P back .

[0096] S506. When there are no noise points and object points in the first grid, add all the points in the first grid to the updated background point set.

[0097] If there are neither noise points nor object points in a certain first grid, add all the points in the first grid to the background point set P back .

[0098] S507. When there are object points but no noise points in the first grid, add the non-object points in the first grid and all the points in the corresponding second grid to the updated background point set to obtain the updated background point cloud.

[0099] If there are object points but no noise points in a certain first grid, add the non-object points in the first grid, and the second depth map Image0 All the points in the corresponding second grid are added to the background point set P back .

[0100] The finally obtained background point set P back is used to update the current background point cloud and generate a new second depth map.

[0101] An embodiment of the present invention also provides a laser point cloud denoising system. Figure 7 is a schematic structural diagram of the laser point cloud denoising system provided by the embodiment of the present invention, as Figure 7 shown, the system includes:

[0102] A preprocessing module 100, configured to obtain the frame point cloud of the current frame, and obtain an initial point cloud after preprocessing the frame point cloud.

[0103] The original point cloud data of the current frame is collected by a lidar, that is, the frame point cloud. All the points in the frame point cloud are in a three-dimensional coordinate system with the lidar as the origin, and each point has three-dimensional coordinates.

[0104] The frame point cloud is an unprocessed point cloud. To speed up the operation speed, preprocessing can be performed before noise point marking to obtain an initial point cloud. The preprocessing includes performing a through filter on the frame point cloud according to the security area, cutting out the effective security area to eliminate data in distant or irrelevant areas; and then removing the point cloud on and below the ground by calculating and fitting the ground result.

[0105] An initial denoising module 200, configured to move the points in the initial point cloud whose reflectivity is lower than the first preset reflectivity threshold and whose distance from the origin is less than the preset distance threshold to the initial noise point set to obtain a processed point cloud; and use the points in the processed point cloud whose reflectivity is lower than the second preset reflectivity threshold and whose distance from the origin is less than the preset distance threshold as the first set of points to be processed; wherein, the second preset reflectivity threshold is greater than the first preset reflectivity threshold.

[0106] The radar will form noise points with significantly lower reflectivity than solid objects in environments such as rain, snow, and fog. However, points with lower reflectivity are not only noise points such as rain, snow, and fog, but may also be caused by the incident angle of the laser or the material of the object. Therefore, it cannot be simply judged directly by the reflectivity value. For example, fog, dust, etc. may also adhere to the object, causing the point cloud of the object to be eroded or generating noise points with slightly higher reflectivity. For the above reasons, it is necessary to conduct a detailed analysis and judgment on the noise points.

[0107] Traverse all the points in the initial point cloud, and the reflectivity is lower than the first preset reflectivity threshold R 1 and the distance from the origin (lidar) is less than the preset distance threshold D 1The discrete points are recorded as the initial noise point set, and these discrete points are removed from the initial point cloud to obtain the processed point cloud. Since severe noise conditions only occur in positions relatively close to the lidar, the low-reflectivity points close to the origin can be directly identified as noise points. Among them, the reflectivity can be directly obtained from the lidar data.

[0108] Then, mark the points with reflectivity lower than the second preset reflectivity threshold R 2 and the distance from the origin less than the preset distance threshold D 1 These points are recorded as the point set to be processed, and the points in the point set to be processed need to be further determined whether they are noise points or attachment points. The second preset reflectivity threshold R 2 is slightly greater than the first preset reflectivity threshold R 1 to obtain all possible noise points.

[0109] The above first preset reflectivity threshold, second preset reflectivity threshold and preset distance threshold can all be manually set according to the actual situation.

[0110] The first noise screening module 300 is used to traverse the first point set to be processed, perform a neighborhood search for each point to be processed in the processed point cloud according to a preset radius to obtain surrounding points, and judge whether the point to be processed is a target noise point according to the distribution characteristics and reflectivity of the surrounding points. If so, determine the target surrounding points corresponding to the target noise point according to the reflectivity, and add the target surrounding points to the first point set to be processed; if not, move the point to be processed as an attachment point from the first point set to be processed to the attachment point set, and update the position and reflectivity of the attachment point in the processed point cloud according to the corresponding surrounding points; after traversal, obtain the second point set to be processed.

[0111] For each point to be processed in the first point set to be processed, use the preset radius D 2 to retrieve and view the distribution characteristics and reflectivity of its surrounding points in the processed point cloud, and perform a preliminary screening of noise points and attachment points according to the distribution characteristics and reflectivity of the surrounding points. For the obtained target noise points, if a point to be processed is determined to be a target noise point, then select the surrounding points suspected of noise among the target surrounding points corresponding to the target noise point according to the reflectivity, that is, the target surrounding points, and add the target surrounding points to the first point set to be processed; otherwise, the point may be an attachment point, add the point to the attachment point set and remove it from the first point set to be processed. Traverse all points in the first point set to be processed, and judge the situation of each point to be processed by the above method to obtain the second point set to be processed and the attachment point set.

[0112] For each point in the attachment point set, it is necessary to restore it to the shape of the object it belongs to, update its position and reflectivity through the distribution characteristics and reflectivity of its surrounding points, and update the updated position result and reflectivity result to the processed point cloud.

[0113] Among them, the preset radius is set manually according to the actual clearing.

[0114] The second noise filtering module 400 is configured to cluster the second point set to be processed to obtain a first clustering result; judge the noise clusters according to the clustering characteristics of the first clustering result, and add the points in the noise clusters to the initial noise point set to obtain the latest noise point set.

[0115] First, construct a depth map (range Image) of the second point set to be processed for the second point set to be processed. If the scanning method of the hardware is petal-shaped scanning, the angles between the points in the scanning result are fixed, so the depth map can be used to ensure that the interval between the scanning points is not uneven due to the change in the distance of the object from the sensor. Of course, if it is a method such as row-by-row scanning, the projection map of the YZ plane can also be used.

[0116] Cluster based on the depth map to obtain a first clustering result, and calculate the clustering characteristics of each cluster in the first clustering result, such as the distribution of points in the cluster and the reflectivity. According to the clustering characteristics, it can be judged whether each cluster is a noise cluster. If a certain cluster is a noise cluster, the points in the cluster are added to the initial noise point set to obtain the latest noise point set. Through the above method, the noise can be further filtered to improve the accuracy of noise marking.

[0117] The noise removal module 500 is configured to perform noise marking on the intrusion point cloud according to the latest noise point set; cluster the intrusion point cloud to obtain a second clustering result, judge the valid clusters and the noise to be removed according to the clustering characteristics of the second clustering result, delete the corresponding noise to be removed from the valid clusters to obtain the target clusters; cancel the marking of the remaining noise in the target clusters, and delete the remaining noise from the latest noise point set, and perform target recognition according to the target clusters; update the current background point cloud according to the processed point cloud and the current background point cloud; wherein, the intrusion point cloud is obtained in advance by differentiating the processed point cloud and the current background point cloud.

[0118] The intrusion point cloud can be obtained after obtaining the processed point cloud. Specifically, it is obtained by background difference, pre-order difference, etc. between the processed point cloud and the current background point cloud to segment the possible intrusion point cloud. Mark the noise in the intrusion point cloud according to the positions of the points in the latest noise point set.

[0119] Cluster the intrusion point cloud to obtain a second clustering result. The clustering can use the Euclidean clustering method based on distance. During the clustering process, it should be noted that the points marked as noise can be added to a certain cluster but not used as the starting point for the next adjacent retrieval to be pushed onto the stack.

[0120] Traverse each cluster in the second clustering result, determine whether it is a valid cluster, and identify the noise points to be removed therein. If it is a valid cluster, delete the noise points to be removed therein to obtain a target cluster without noise points. Cancel the marking of the remaining noise points within the target cluster and delete the remaining noise points from the latest noise point set. For the obtained target cluster, subsequent target recognition processes can be performed to obtain recognition results. The target recognition process can be completed using a recognition model, and the target cluster can be updated after recognition.

[0121] Furthermore, background point clouds are required in the differential process, and the accuracy of the background point clouds will also affect the final recognition accuracy. For example, relatively thick fog or dust sometimes causes certain occlusion of the background, resulting in the background being re-exposed after the fog or dust moves and being misrecognized as an intrusion object. Therefore, a real-time updated background needs to be constructed to assist in the differential. Combine the processed point cloud with the current background point cloud according to the noise point distribution in the processed point cloud to obtain an updated background point cloud for use in the differential process of the next frame.

[0122] As an alternative implementation Figure 8 is a schematic structural diagram of the first noise screening module for target noise determination provided by an embodiment of the present invention. As Figure 8 shown, the first noise screening module 300 includes:

[0123] The first target noise determination sub-module 3001 is used to retrieve the surrounding points with a reflectivity higher than or equal to the second preset reflectivity threshold, denoted as the first surrounding points, count the number and positions of the first surrounding points, and when the number of the first surrounding points is 0, regard the corresponding point to be processed as a target noise point.

[0124] For each point to be processed, count the surrounding points with a reflectivity greater than or equal to the second preset reflectivity threshold R 2 among its corresponding surrounding points, regard them as the first surrounding points, and count the number and positions of the first surrounding points. If the number of the first surrounding points is 0, or the number of the first surrounding points is lower than a certain threshold, then it is considered that this point must be a noise point, denoted as a target noise point.

[0125] The second target noise determination sub-module 3002 is used to retrieve the surrounding points with a reflectivity lower than the second preset reflectivity threshold, denoted as the second surrounding points, count the number and positions of the second surrounding points, calculate the distribution density of the second surrounding points according to the number of the second surrounding points, and regard the point to be processed corresponding to the distribution density lower than the first preset density threshold as a target noise point.

[0126] For each point to be processed, count the surrounding points with a reflectivity less than the second preset reflectivity threshold R 2The peripheral points are taken as the second peripheral points, and the number and positions of the second peripheral points are counted. According to the number of the second peripheral points, the distribution density of the low reflectivity points around the point to be processed can be calculated. If the corresponding distribution density of the point to be processed is lower than the first preset density threshold T 1 , it is considered that this point must be a noise point, denoted as the target noise point.

[0127] The third target noise point determination sub-module 3003 is used to calculate the first reflectivity mean value and the first standard deviation of the reflectivities of all peripheral points corresponding to each point to be processed, and take the point to be processed corresponding to the first standard deviation lower than the first preset standard deviation threshold as the target noise point.

[0128] For each point to be processed, calculate the first reflectivity mean value of its peripheral points and the first standard deviation of the reflectivities. If the first standard deviation is lower than the first preset standard deviation threshold, it is considered that this point to be processed must be a noise point, denoted as the target noise point.

[0129] As an optional implementation manner, the first preset standard deviation threshold is usually 0.5×R 2 .

[0130] For a point to be processed, as long as one of the above three judgment conditions is met, it is considered to be the target noise point.

[0131] The target peripheral point determination sub-module 3004 is used to take the peripheral points with reflectivity lower than the first reflectivity mean value as the target peripheral points corresponding to the target noise points.

[0132] If a certain point to be processed is the target noise point, then take the peripheral points with reflectivity less than the first reflectivity mean value among its peripheral points as the target peripheral points and add them to the first set of points to be processed.

[0133] As an optional implementation manner, Figure 9 is the structural schematic diagram of the first noise screening module for attachment point update provided by the embodiment of the present invention. As shown in Figure 9 , the first noise screening module 300 further includes:

[0134] The reflectivity update sub-module 3005 is used to calculate the second reflectivity mean value of the first peripheral points corresponding to each attachment point, and take the second reflectivity mean value as the updated reflectivity of the corresponding attachment point;

[0135] For each attachment point in the attachment point set, calculate the reflectivity mean value of its corresponding first peripheral points, that is, the second reflectivity mean value, as the new reflectivity of this attachment point.

[0136] The reflectivity ratio calculation sub-module 3006 is used to calculate the quantity ratio of the first peripheral points to the second peripheral points, and calculate the reflectivity ratio of the updated reflectivity to the unupdated reflectivity;

[0137] For each attachment point, calculate the ratio r of the number of first peripheral points to the number of second peripheral points 附着1 , and calculate the reflectivity ratio r of the updated reflectivity to the original reflectivity 附着2 .

[0138] A central position calculation sub-module 3007 for calculating the first central position of the first peripheral points and the second central position of the second peripheral points;

[0139] Calculate the first central position c of the first peripheral points respectively 附着1 and the second central position c of the second peripheral points 附着1 . Since the positions of each first peripheral point and each second peripheral point have been counted before, the corresponding central positions can be directly calculated through the position coordinates.

[0140] A position update sub-module 3008 for calculating the updated position according to the quantity ratio, the reflectivity ratio, the first central position and the second central position.

[0141] As an optional implementation manner, calculate the updated position c according to the quantity ratio, the reflectivity ratio, the first central position and the second central position 附着 by the following formula:

[0142] c 附着 =c 附着1 +(c 附着2 -c 附着1 )×(r 附着1 ×w 1 +r 附着2 ×w 2 ). Where w 1 and w 2 are specified weights

[0143] As an optional implementation manner, Figure 10 is a schematic structural diagram of the second noise filtering module provided by the embodiments of the present invention, as Figure 10 shown, the second noise filtering module 400 includes:

[0144] A calculation sub-module 4001 for calculating the length of each cluster in the first clustering result in depth, calculating the third reflectivity mean value and the second standard deviation of all points within the cluster;

[0145] For each cluster in the first clustering result, calculate its length in depth, calculate the third reflectivity mean value and the second standard deviation of the inner points thereof, count the number of inner points in the cluster, and calculate the point density within the cluster based on the quantity.

[0146] The first noise cluster determination sub-module 4002 is configured to use, as a noise cluster, a cluster whose length is less than a first preset length threshold, whose third reflectance mean is less than a first preset reflectance threshold, and whose second standard deviation is less than a second preset standard deviation threshold;

[0147] For each cluster, if its corresponding length is less than a first preset length threshold L 1 , and at the same time its third reflectance mean is less than a first preset reflectance threshold R 1 , and at the same time its second standard deviation is less than a second preset standard deviation threshold (usually 0.5×R 1 ), then it is considered that this cluster is a noise cluster, that is, the points therein must be noise points, and all the points within the cluster are added to the initial noise point set.

[0148] The second noise cluster determination sub-module 4003 is configured to use, as a noise cluster, a cluster whose number of points is less than a preset number threshold or whose point density is less than a second preset density threshold;

[0149] For each cluster, if the number of points within the cluster is less than a preset number threshold, or if the point density is less than a second preset density threshold, it is considered that the points within this cluster may be noise points but are relatively discrete and may not affect subsequent detection. The cluster is used as a noise cluster, and the points within the cluster are added to the initial noise point set.

[0150] The third noise cluster determination sub-module 4004 is configured to thin out a cluster and then calculate the Gaussian curvature, calculate the proportion of the first number of points with a Gaussian curvature lower than a preset curvature threshold in the cluster to which they belong, and when the first number proportion is greater than a first preset proportion threshold, use the corresponding cluster as a noise cluster.

[0151] For the remaining clusters after the above two judgment methods, they may be noise clusters and need to be further judged. First, thin out the points within the cluster, calculate the Gaussian curvature of the thinned points, count the number of points with a Gaussian curvature lower than the preset curvature threshold, and calculate the proportion of the first number of these points to the total number of points within the cluster to which they belong. If the proportion is too high (greater than a first preset proportion threshold), then it is considered that this cluster is a noise cluster, and the points therein are added to the initial noise point set; otherwise, this cluster is not a noise cluster.

[0152] As an alternative implementation manner, Figure 11 is a schematic structural diagram of a noise removal module for denoising provided by an embodiment of the present invention. As Figure 11 shown, the noise removal module 500 includes:

[0153] The first denoising sub-module 5001 is used to count the number of noise points and the proportion of noise points in each cluster in the second clustering result; clusters with the number of noise points lower than the preset noise point number threshold, or clusters with the proportion of noise points lower than the preset noise point proportion threshold are regarded as valid clusters, and the noise points in the valid clusters are regarded as noise points to be removed.

[0154] For each cluster in the second clustering result, if there are no noise points in the cluster (the number of noise points is lower than the preset noise point number threshold) or the proportion of noise points is small (the proportion of noise points is lower than the preset noise point proportion threshold), the cluster is considered a valid cluster, and all the marked noise points in the cluster are the noise points to be removed. Remove these noise points to be removed from the valid cluster.

[0155] The valid cluster determination sub-module 5002 is used to calculate the point distance from each noise point in the cluster to the nearest non-noise point, calculate the second quantity proportion of the points with the point distance greater than or equal to the second preset length threshold in the cluster to which they belong. When the second quantity proportion is greater than the second preset proportion threshold, the corresponding cluster is regarded as an invalid cluster, otherwise it is a valid cluster;

[0156] For the remaining clusters, calculate the point distance from each noise point to the nearest non-noise point, count the number of points with the point distance greater than or equal to the second preset length threshold D 3 and calculate its second quantity proportion to the total number of points in the cluster. If the second quantity proportion is relatively high (the second quantity proportion is greater than the second preset proportion threshold), then the cluster is considered invalid, otherwise it is valid.

[0157] The second denoising sub-module 5003 is used to regard the points with a point distance less than the second preset length threshold in the valid cluster as points to be updated, calculate the position and reflectivity to be updated according to the surrounding points corresponding to the points to be updated, and judge whether there are points with a distance exceeding the second preset length threshold within the preset radius of the position to be updated. If not, update the position and reflectivity of the points to be updated with the position and reflectivity to be updated respectively; regard the points to be updated that have not been updated as noise points to be removed.

[0158] For the valid clusters determined to be valid, regard the points with a point distance less than the second preset length threshold D 3 as attachment points (points to be updated). The attachment points need to be updated. Calculate the position to be updated and the reflectivity to be updated according to the distribution and reflectivity of the surrounding points of the points to be updated. The specific update method can be the same as the update method of the above attachment points.

[0159] Judge whether there are points with a distance greater than the second preset length threshold D within the preset radius D 2 near the updated position 3If the points do not exist, update using the position to be updated and the reflectivity to be updated; otherwise, do not update. Use the points to be updated that are not updated as the points to be denoised, and remove all the points to be denoised within the effective cluster. Cancel the denoising marks for the remaining noise points (updated noise points) within the effective cluster, and delete these points from the latest set of noise points.

[0160] As an alternative implementation Figure 12 is a schematic structural diagram of the noise removal module for background reconstruction provided by an embodiment of the present invention. As Figure 12 shown, the noise removal module 500 includes:

[0161] An update judgment sub-module 5004, which is used to count the proportion of noise points within the first preset distance threshold range of the points in the processed point cloud. When the proportion of noise points is greater than the third preset proportion threshold, do not update the current background point cloud; otherwise, construct the first depth map of the initial point cloud and the second depth map of the current background point cloud, and perform grid division on the first depth map and the second depth map.

[0162] When attempting to construct the background each time, count the proportion of noise points within the first preset distance threshold range of the processed point cloud. If the proportion is too large (the proportion of noise points is greater than the third preset proportion threshold), give up updating the current background point cloud; otherwise, perform the update.

[0163] During the update, first construct the first depth map Image 1 of the processed point cloud, and at the same time construct the second depth map Image 0 of the current background point cloud. Perform grid division on the first depth map and the second depth map according to the same rules.

[0164] The first grid processing sub-module 5005 is used for each first grid in the first depth map. When there are noise points in it, add all the points in the second grid corresponding to the first grid in the second depth map to the updated background point set, and add all the points in the eight grids around the corresponding second grid to the updated background point set.

[0165] For each first grid in the first depth map, when there are noise points in a certain first grid, add all the points in the second grid corresponding to the first grid in the second depth map Image 0 to the updated background point set P back , and consider that all the eight grids near the second grid are added to the background point set P back .

[0166] The second grid processing sub-module 5006 is used to add all the points in the first grid to the updated background point set when there are no noise points and object points in the first grid.

[0167] If there are neither noise points nor object points in a certain first grid, then add all the points in the first grid to the background point set P back to it.

[0168] The third grid processing sub-module 5007 is used to, when there are object points but no noise points in the first grid, add the non-object points in the first grid and all the points in the corresponding second grid to the updated background point set to obtain the updated background point cloud.

[0169] If there are object points but no noise points in a certain first grid, then add the non-object points in the first grid and the points in the corresponding second grid in the second depth map Image 0 to the background point set P in its entirety. back to it.

[0170] The finally obtained background point set P back is used to update the current background point cloud and generate a new second depth map.

[0171] The above technical solution has the following beneficial effects: By analyzing information such as the distribution of points and reflectivity in the initial point cloud, noise points are gradually filtered, and possible noise points are processed one by one according to their causes to improve the accuracy of removing noise caused by extreme weather; The noise points at the edges of objects are screened out and restored to the contour of the object itself as attachments, which is convenient for subsequent target recognition and improves the accuracy of subsequent target recognition; The background is selectively updated, and the true situation of the background is restored according to the screened noise points, thereby ensuring the accuracy of the intrusion point cloud after background difference and improving the accuracy of subsequent target recognition; The traditional algorithm is used to improve the operation speed and ensure real-time performance.

[0172] The specific implementation manners of the above invention further elaborate on the purpose, technical solution and beneficial effects of the present invention. It should be understood that the above content is only the specific implementation manners of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A laser point cloud denoising method, characterized in that: include: Obtaining a frame point cloud of the current frame, and obtaining an initial point cloud after preprocessing the frame point cloud; From the initial point cloud, points whose reflectivity is lower than a first preset reflectivity threshold and whose distance from the origin is less than a preset distance threshold are moved to the initial noise point set to obtain a processed point cloud; points in the processed point cloud whose reflectivity is lower than a second preset reflectivity threshold and whose distance from the origin is less than the preset distance threshold are taken as a first point set to be processed; wherein the second preset reflectivity threshold is greater than the first preset reflectivity threshold; Traversing the first set of points to be processed, performing a neighborhood search in the processing point cloud according to a preset radius for each point to be processed to obtain surrounding points, judging whether the point to be processed is a target noise point according to the distribution characteristics and reflectivity of the surrounding points, and if so, determining the target surrounding points corresponding to the target noise point according to the reflectivity, and adding the target surrounding points to the first set of points to be processed; if not, moving the point to be processed from the first set of points to be processed to the attachment point set as an attachment point, and updating the position and reflectivity of the attachment point in the processing point cloud according to the corresponding surrounding points; after traversing, obtaining a second set of points to be processed; Clustering the second set of points to be processed to obtain a first clustering result; determining a noise point cluster according to the clustering characteristics of the first clustering result, and adding points in the noise point cluster to the initial noise point set to obtain a latest noise point set; The intrusion point cloud is marked with noise points according to the latest noise point set; the intrusion point cloud is clustered to obtain a second clustering result, and effective clusters and noise points to be removed are determined according to the clustering characteristics of the second clustering result, and the corresponding noise points to be removed are deleted from the effective clusters to obtain a target cluster; the marking of the remaining noise points in the target cluster is canceled, and the remaining noise points are deleted from the latest noise point set, and the target is identified according to the target cluster; the current background point cloud is updated according to the processed point cloud and the current background point cloud; wherein the intrusion point cloud is obtained in advance by differentially comparing the processed point cloud with the current background point cloud.

2. The laser point cloud denoising method according to claim 1, characterized in that: The step of judging whether the point to be processed is a target noise point according to the distribution characteristics and reflectivity of the surrounding points, and if so, determining the target surrounding points corresponding to the target noise point according to the reflectivity, comprises: Retrieving peripheral points whose reflectivity is higher than or equal to a second preset reflectivity threshold, recording them as first peripheral points, counting the number and positions of the first peripheral points, and when the number of the first peripheral points is 0, taking the corresponding point to be processed as a target noise point; Retrieve surrounding points whose reflectivity is lower than a second preset reflectivity threshold, record them as second surrounding points, count the number and positions of the second surrounding points, calculate the distribution density of the second surrounding points according to the number of the second surrounding points, and take the points to be processed corresponding to the distribution density lower than the first preset density threshold as target noise points; Calculate the first reflectivity mean and the first standard deviation of the reflectivity of all surrounding points corresponding to each to-be-processed point, and take the to-be-processed point corresponding to the first standard deviation lower than the first preset standard deviation threshold as the target noise point; The peripheral points whose reflectivity is lower than the first reflectivity average are used as target peripheral points corresponding to the target noise points.

3. The laser point cloud denoising method according to claim 2, characterized in that: The updating of the position and reflectivity of the attachment point in the processed point cloud according to the corresponding surrounding points includes: Calculate the second reflectivity average of the first peripheral points corresponding to each attachment point, and use the second reflectivity average as the updated reflectivity of the corresponding attachment point; Calculating a ratio of the number of the first peripheral points to the number of the second peripheral points, and calculating a reflectivity ratio of the updated reflectivity to the unupdated reflectivity; Calculate a first center position of a first peripheral point and a second center position of a second peripheral point; An updated position is calculated according to the quantity ratio, the reflectivity ratio, the first center position and the second center position.

4. The laser point cloud denoising method according to claim 1, characterized in that: The determining the noise point clustering according to the clustering feature of the first clustering result includes: Calculate the length of each cluster in the first clustering result in depth, and calculate the third reflectivity mean and second standard deviation of all points in the cluster; The clusters whose length is less than the first preset length threshold, whose third reflectivity mean is less than the first preset reflectivity threshold, and whose second standard deviation is less than the second preset standard deviation threshold are regarded as noise clusters; The clusters whose number of points is less than a preset number threshold, or the clusters whose point density is less than a second preset density threshold, are regarded as noise clusters; After thinning the clusters, the Gaussian curvature is calculated, and the first quantity ratio of points whose Gaussian curvature is lower than the preset curvature threshold in the corresponding cluster is calculated. When the first quantity ratio is greater than the first preset ratio threshold, the corresponding cluster is regarded as a noise cluster.

5. The laser point cloud denoising method according to claim 1, characterized in that: The determining effective clusters and noise points to be removed according to the clustering features of the second clustering result includes: Counting the number of noise points and the proportion of noise points in each cluster in the second clustering result; taking clusters whose number of noise points is lower than a preset noise point number threshold, or taking clusters whose proportion of noise points is lower than a preset noise point number proportion threshold as valid clusters, and taking noise points in the valid clusters as noise points to be removed; Calculate the point distance from each noise point to the nearest non-noise point in the cluster, calculate the second quantity ratio of points whose point distance is greater than or equal to the second preset length threshold in the corresponding cluster, and when the second quantity ratio is greater than the second preset ratio threshold, regard the corresponding cluster as an invalid cluster, otherwise it is a valid cluster; The points whose midpoint distance in the valid cluster is less than the second preset length threshold are taken as the points to be updated, and the position and reflectivity to be updated are calculated according to the surrounding points corresponding to the points to be updated. It is determined whether there are points within the preset radius of the position to be updated whose distance exceeds the second preset length threshold. If not, the position and reflectivity to be updated are used to update the position and reflectivity of the points to be updated respectively; the points to be updated that have not been updated are taken as the noise points to be removed.

6. The laser point cloud denoising method according to claim 1, characterized in that: The updating of the current background point cloud according to the processed point cloud and the current background point cloud comprises: Counting the proportion of noise points in the points in the processed point cloud within the first preset distance threshold range, when the proportion of noise points is greater than a third preset proportion threshold, not updating the current background point cloud; otherwise, constructing a first depth map of the initial point cloud and a second depth map of the current background point cloud, and performing grid division on the first depth map and the second depth map; For each first grid in the first depth map, when there is a noise point therein, all points in the second grid corresponding to the first grid in the second depth map are added to the updated background point set, and all points in the eight grids surrounding the corresponding second grid are added to the updated background point set; When there are no noise points or object points in the first grid, all points in the first grid are added to the updated background point set; When there are object points but no noise points in the first grid, the non-object points in the first grid and all points in the corresponding second grid are added to the updated background point set to obtain an updated background point cloud.

7. A laser point cloud denoising system, characterized in that: include: A preprocessing module is used to obtain a frame point cloud of a current frame, and obtain an initial point cloud after preprocessing the frame point cloud; The initial denoising module is used to move points whose reflectivity is lower than a first preset reflectivity threshold and whose distance from the origin is less than a preset distance threshold from the initial point cloud to the initial noise point set to obtain a processed point cloud; and to take points whose reflectivity is lower than a second preset reflectivity threshold and whose distance from the origin is less than the preset distance threshold in the processed point cloud as a first point set to be processed; wherein the second preset reflectivity threshold is greater than the first preset reflectivity threshold; A first noise point screening module is used to traverse the first set of points to be processed, perform a neighborhood search in the processing point cloud according to a preset radius for each point to be processed to obtain surrounding points, determine whether the point to be processed is a target noise point according to the distribution characteristics and reflectivity of the surrounding points, and if so, determine the target surrounding points corresponding to the target noise point according to the reflectivity, and add the target surrounding points to the first set of points to be processed; if not, move the point to be processed from the first set of points to be processed to the attachment point set as an attachment point, and update the position and reflectivity of the attachment point in the processing point cloud according to the corresponding surrounding points; after traversal, a second set of points to be processed is obtained; A second noise point screening module is used to cluster the second to-be-processed point set to obtain a first clustering result; determine the noise point cluster according to the clustering characteristics of the first clustering result, and add the points in the noise point cluster to the initial noise point set to obtain the latest noise point set; A noise removal module is used to mark the noise points of the intrusion point cloud according to the latest noise point set; cluster the intrusion point cloud to obtain a second clustering result, determine the effective clustering and the noise points to be removed according to the clustering characteristics of the second clustering result, delete the corresponding noise points to be removed from the effective clustering to obtain the target cluster; cancel the marking of the remaining noise points in the target cluster, and delete the remaining noise points from the latest noise point set, and perform target identification according to the target cluster; update the current background point cloud according to the processed point cloud and the current background point cloud; wherein the intrusion point cloud is obtained in advance by differentially analyzing the processed point cloud and the current background point cloud.

8. The laser point cloud denoising system according to claim 7, characterized in that: The first noise point screening module includes: A first target noise point determination submodule is used to retrieve surrounding points whose reflectivity is higher than or equal to a second preset reflectivity threshold value, record them as first surrounding points, count the number and positions of the first surrounding points, and when the number of the first surrounding points is 0, take the corresponding point to be processed as a target noise point; The second target noise point determination submodule is used to retrieve surrounding points whose reflectivity is lower than a second preset reflectivity threshold value, record them as second surrounding points, count the number and positions of the second surrounding points, calculate the distribution density of the second surrounding points according to the number of the second surrounding points, and take the to-be-processed points corresponding to the distribution density lower than the first preset density threshold value as target noise points; A third target noise point determination submodule is used to calculate a first reflectivity mean and a first standard deviation of the reflectivity of all surrounding points corresponding to each to-be-processed point, and to take the to-be-processed point corresponding to the first standard deviation lower than a first preset standard deviation threshold as a target noise point; The target peripheral point determination submodule is used to take peripheral points whose reflectivity is lower than the first reflectivity average as target peripheral points corresponding to the target noise points.

9. The laser point cloud denoising system according to claim 7, characterized in that: The second noise point screening module includes: A calculation submodule, used to calculate the length of each cluster in the first clustering result in depth, and calculate the third reflectivity mean and second standard deviation of all points in the cluster; A first noise point cluster determination submodule, configured to take a cluster whose length is less than a first preset length threshold, whose third reflectivity mean is less than the first preset reflectivity threshold, and whose second standard deviation is less than the second preset standard deviation threshold as a noise point cluster; A second noise point cluster determination submodule, used to take a cluster whose number of points is less than a preset number threshold, or a cluster whose point density is less than a second preset density threshold, as a noise point cluster; The third noise point cluster determination submodule is used to calculate the Gaussian curvature after thinning the clusters, and calculate the first quantity ratio of points whose Gaussian curvature is lower than the preset curvature threshold in the corresponding cluster. When the first quantity ratio is greater than the first preset ratio threshold, the corresponding cluster is regarded as a noise point cluster.

10. The laser point cloud denoising system according to claim 7, characterized in that: The noise removal module comprises: A first denoising submodule is used to count the number of noise points and the proportion of the number of noise points in each cluster in the second clustering result; a cluster whose number of noise points is lower than a preset noise point number threshold, or a cluster whose proportion of the number of noise points is lower than a preset noise point number threshold is taken as a valid cluster, and the noise points in the valid cluster are taken as noise points to be removed; The effective cluster determination submodule is used to calculate the point distance from each noise point to the nearest non-noise point in the cluster, and calculate the second quantity ratio of points whose point distance is greater than or equal to the second preset length threshold in the corresponding cluster. When the second quantity ratio is greater than the second preset ratio threshold, the corresponding cluster is regarded as an invalid cluster, otherwise it is a valid cluster; The second denoising submodule is used to take the points whose midpoint distance in the effective cluster is less than the second preset length threshold as the points to be updated, calculate the position and reflectivity to be updated according to the surrounding points corresponding to the points to be updated, and determine whether there are points whose distance exceeds the second preset length threshold within the preset radius of the position to be updated. If not, the position and reflectivity to be updated are used to update the position and reflectivity of the point to be updated respectively; and the points to be updated that have not been updated are taken as the noise points to be removed.

Images