Laser point cloud registration method and system in automatic driving

Abstract

The invention relates to a laser point cloud registration method and system in automatic driving, and the method comprises the steps: obtaining a source point cloud and a target point cloud shot by a laser radar in automatic driving, initializing a noise weight based on the density, and calculating a fast point feature histogram of the source point cloud and the target point cloud; performing down-sampling on the source point cloud and the target point cloud according to the fast point feature histogram and the number of iterations, calculating a matching probability from a source point in the source point cloud to a target point of the target point cloud, and counting the source point of which the matching probability is lower than a threshold value as a noise point; calculating the noise weight of the next iteration by adopting a dynamic noise estimation mode according to the number of the noise points; optimizing the transformation matrix and the translation vector based on the matching probability, judging the variable quantity of the transformation matrix and the translation vector relative to the transformation matrix and the translation vector calculated last time, and stopping updating the transformation matrix and the translation vector when the variable quantity meets a preset condition; and transforming the source point cloud by using the transformation matrix and the translation vector to complete registration of the source point cloud and the target point cloud.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving, and in particular to a laser point cloud registration method and system for autonomous driving. Background Art

[0002] Autonomous driving technology can not only improve traffic efficiency and reduce traffic accidents, but also free the driver's hands and improve the comfort and convenience of travel. With the continuous development of autonomous driving technology, how to enable the autonomous driving system to accurately perceive the environment and make decisions in a complex and changing environment has become one of the core issues of technical implementation. In this regard, laser radar (LIDAR) has become a key tool for obtaining environmental information in autonomous driving systems due to its high-precision and long-distance detection capabilities. LiDAR can generate high-density three-dimensional point cloud data, reflecting the geometric shape and spatial distribution of surrounding objects, helping autonomous driving systems complete perception tasks such as obstacle detection and path planning.

[0003] In autonomous driving, point cloud registration technology is one of the core technologies for environmental perception and map updating. The task of point cloud registration is to align point cloud data from different perspectives, different times or different sensors to generate a unified 3D model or scene graph. The scanning results of LiDAR in different environments and conditions often contain different types of noise, such as sensor errors, interference from dynamic objects, inconsistency of environmental reflections, etc. These noises will make the results of point cloud registration inaccurate, which will affect the reliability and safety of the autonomous driving system. Summary of the invention

[0004] Aiming at the problem that point cloud registration in autonomous driving is easily affected by noise, resulting in inaccurate modeling or obstacle recognition, the present invention first provides a laser point cloud registration method in autonomous driving, the method comprising the following steps: Obtain the source point cloud and target point cloud taken by the LiDAR in autonomous driving, initialize the noise weight based on the density of the source point cloud and target point cloud, and calculate the fast point feature histogram of the source point cloud and target point cloud; The source point cloud and the target point cloud are downsampled according to the fast point feature histogram and the number of iterations, the matching probability of the source point in the source point cloud to the target point in the target point cloud is calculated, the source points with matching probability lower than the threshold are counted as noise points, and the noise weight of the next iteration is calculated by dynamic noise estimation according to the number of noise points; Based on the matching probability, the transformation matrix and translation vector are optimized, and the change of the transformation matrix and translation vector relative to the last calculated transformation matrix and translation vector is determined. When the change meets the preset conditions, the transformation matrix and translation vector are stopped from being updated. The source point cloud is transformed using the transformation matrix and translation vector to complete the registration of the source point cloud and the target point cloud.

[0005] Preferably, the noise weight is initialized based on the density of the source point cloud and the target point cloud, specifically: For each point in the source point cloud and the target point cloud, calculate the density of points in its neighborhood; After calculating the local density of each point in the source point cloud and the target point cloud, the density distribution is obtained, the mean of the density is obtained, and the initialization noise weight is obtained using the mean of the density.

[0006] Preferably, downsampling the source point cloud and the target point cloud according to the fast point feature histogram and the number of iterations is specifically: Evaluate the importance score of each point based on its fast point feature histogram feature; Sort the source point cloud and the target point cloud respectively according to the importance scores of the points, obtain the number of iterations, and determine the proportion of points retained in this iteration according to the number of iterations; According to the ratio, the point with the highest importance score is selected from the sorted source point cloud as the point of the source point cloud, and according to the ratio, the point with the highest importance score is selected from the sorted target point cloud as the point of the target point cloud.

[0007] Preferably, the noise weight of the next iteration is calculated by adopting a dynamic noise estimation method according to the number of noise points, specifically: Calculate the ratio of the number of noise points to the number of points in the source point cloud in this iteration; The neighborhood of the source point in the original source point cloud is obtained, and the normal difference between the source point and the neighborhood is calculated, and an adjustment coefficient is obtained according to the normal difference. The noise weight of the source point in the next iterative calculation is obtained by the adjustment coefficient and the ratio.

[0008] Preferably, the adjustment coefficient is obtained according to the normal difference, specifically: Calculate the average of the dot products of the unit normal vectors of the source point and each point in the neighborhood; The adjustment factor is 1 minus the average value.

[0009] Preferably, the importance score of the point is evaluated based on the fast point feature histogram feature of each point, specifically: Calculate the similarity of the fast point feature histograms of the source point and the target point. For each point in the source point cloud and the target point cloud, take the maximum similarity as the score of the point. The local density of the point in the original source point cloud or the original target point cloud is obtained, and the weighted value is obtained by using the local density. The score is weighted by using the weighted value to obtain the importance score of the point.

[0010] Secondly, the present invention also provides a laser point cloud registration system for autonomous driving, the system comprising the following modules: The initialization module is used to obtain the source point cloud and target point cloud captured by the lidar in autonomous driving, initialize the noise weight based on the density of the source point cloud and the target point cloud, and calculate the fast point feature histogram of the source point cloud and the target point cloud; The downsampling module is used to downsample the source point cloud and the target point cloud according to the fast point feature histogram and the number of iterations, calculate the matching probability of the source point in the source point cloud to the target point in the target point cloud, count the source points with matching probability lower than the threshold as noise points, and calculate the noise weight of the next iteration by dynamic noise estimation according to the number of noise points; The registration module is used to optimize the transformation matrix and translation vector based on the matching probability, determine the change of the transformation matrix and translation vector relative to the last calculated transformation matrix and translation vector, and stop updating the transformation matrix and translation vector when the change meets the preset conditions; use the transformation matrix and translation vector to transform the source point cloud to complete the registration of the source point cloud and the target point cloud.

[0011] Preferably, the noise weight is initialized based on the density of the source point cloud and the target point cloud, specifically: For each point in the source point cloud and the target point cloud, calculate the density of points in its neighborhood; After calculating the local density of each point in the source point cloud and the target point cloud, the density distribution is obtained, the mean of the density is obtained, and the initialization noise weight is obtained using the mean of the density.

[0012] Preferably, downsampling the source point cloud and the target point cloud according to the fast point feature histogram and the number of iterations is specifically: Evaluate the importance score of each point based on its fast point feature histogram feature; Sort the source point cloud and the target point cloud respectively according to the importance scores of the points, obtain the number of iterations, and determine the proportion of points retained in this iteration according to the number of iterations; According to the ratio, the point with the highest importance score is selected from the sorted source point cloud as the point of the source point cloud, and according to the ratio, the point with the highest importance score is selected from the sorted target point cloud as the point of the target point cloud.

[0013] Preferably, the noise weight of the next iteration is calculated by adopting a dynamic noise estimation method according to the number of noise points, specifically: Calculate the ratio of the number of noise points to the number of points in the source point cloud in this iteration; The neighborhood of the source point in the original source point cloud is obtained, and the normal difference between the source point and the neighborhood is calculated, and an adjustment coefficient is obtained according to the normal difference. The noise weight of the source point in the next iterative calculation is obtained by the adjustment coefficient and the ratio.

[0014] Preferably, the adjustment coefficient is obtained according to the normal difference, specifically: Calculate the average of the dot products of the unit normal vectors of the source point and each point in the neighborhood; The adjustment factor is 1 minus the average value.

[0015] Preferably, the importance score of the point is evaluated based on the fast point feature histogram feature of each point, specifically: Calculate the similarity of the fast point feature histograms of the source point and the target point. For each point in the source point cloud and the target point cloud, take the maximum similarity as the score of the point. The local density of the point in the original source point cloud or the original target point cloud is obtained, and the weighted value is obtained by using the local density. The score is weighted by using the weighted value to obtain the importance score of the point.

[0016] In the initial stage of the present invention, points in dense areas are assigned lower noise weights, which can more accurately distinguish noise points from important points and improve the robustness of registration. Moreover, in the CPD iteration, combined with the matching probability and FPFH, the noise weight can be adjusted more flexibly to adapt to the dynamic changes of noise points in the point cloud registration process, making the registration result more accurate and stable, and suitable for complex autonomous driving scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of Embodiment 1; Figure 2 This is a structural diagram of the second embodiment. DETAILED DESCRIPTION

[0018] In this article, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings 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.

[0020] In a first embodiment of the present invention, a laser point cloud registration method for autonomous driving is provided. Figure 1 As shown, the method comprises the following steps: S1, obtain the source point cloud and target point cloud taken by the laser radar in autonomous driving, initialize the noise weight based on the density of the source point cloud and the target point cloud, and calculate the fast point feature histogram of the source point cloud and the target point cloud; In autonomous driving, the LiDAR continuously acquires point clouds around the vehicle. When the point cloud at the previous moment is aligned with the point cloud at the next moment, the point cloud at the previous moment is the target point cloud and the point cloud at the next moment is the source point cloud. Alternatively, among the point clouds acquired at the same moment, one point cloud acquired by different LiDAR sensors is the source point cloud and the other is the target point cloud.

[0021] The Coherent Point Drift (CPD) algorithm includes the E step and the M step. In the E step, the formula is used. Calculation Point The probability distribution of Represents the noise weight. When the noise weight is small, the noise points may be forced to participate in the matching, resulting in mismatching; when the weight of the noise point is large, a large number of points are regarded as potential noise, which may weaken the geometric constraints on the real points and reduce the registration accuracy.

[0022] At the beginning of the registration, the noise weight is initialized based on the density of the point cloud (i.e., the distance between points or the distribution of points). For example, if the density of the source point cloud and the target point cloud is low, it may mean that there are more noise points, so these points will be given a higher noise weight during initialization. In an optional embodiment, the noise weight is initialized based on the density of the source point cloud and the target point cloud, specifically: For each point in the source point cloud and the target point cloud, calculate the density of points in its neighborhood; After calculating the local density of each point in the source point cloud and the target point cloud, the density distribution is obtained, the mean of the density is obtained, and the initialization noise weight is obtained using the mean of the density.

[0023] For each point in the source point cloud and the target point cloud, the density of the points in its neighborhood is calculated. Density usually refers to the number of points within a certain radius around a point. The local density distribution of each point is calculated based on these density values, and the density mean of all points in the source point cloud and the target point cloud is calculated. The average of the density is used to obtain the initialization noise weight. Specifically, the initialization noise weight is calculated using a negative exponential function or an inverse proportional function. If the point cloud is relatively sparse, then the density is low and the calculated noise weight is higher, which means that the point may contain more noise, otherwise the noise weight is lower. In one embodiment, the initialization noise weights of all points are the same.

[0024] Fast Point Feature Histogram (FPFH) is a method for describing the local geometric features of each point in a point cloud, and is particularly suitable for identifying matching points in point cloud registration. The calculation process of fast point feature histogram belongs to the prior art and will not be described here.

[0025] S2, downsample the source point cloud and the target point cloud according to the fast point feature histogram and the number of iterations, calculate the matching probability of the source point in the source point cloud to the target point in the target point cloud, count the source points with matching probability lower than the threshold as noise points, and calculate the noise weight of the next iteration by dynamic noise estimation according to the number of noise points; In the registration process, the FPFH feature can better match the points in the source point cloud and the target point cloud. In order to speed up the calculation and avoid processing too much data, the point cloud is downsampled according to the FPFH and the number of iterations, that is, the number of points in the point cloud is reduced. In one embodiment, the source point cloud and the target point cloud are downsampled according to the fast point feature histogram and the number of iterations, specifically: Evaluate the importance score of each point based on its fast point feature histogram feature; Sort the source point cloud and the target point cloud respectively according to the importance scores of the points, obtain the number of iterations, and determine the proportion of points retained in this iteration according to the number of iterations; According to the ratio, the point with the highest importance score is selected from the sorted source point cloud as the point of the source point cloud, and according to the ratio, the point with the highest importance score is selected from the sorted target point cloud as the point of the target point cloud.

[0026] Calculate the source point cloud and the target point cloud for each point In step S1, the fast point feature histogram of each point in the source point cloud and the target point cloud is calculated, and the importance score of the point is evaluated according to the FPFH of each point in the source point cloud and the target point cloud. In a specific embodiment, the similarity of the fast point feature histograms of the source point and the target point is calculated, and for each point in the source point cloud and the target point cloud, the maximum similarity is used as the score of the point; the local density of the point in the original source point cloud or the original target point cloud is obtained, and the weighted value is obtained by using the local density, and the score is weighted by the weighted value to obtain the importance score of the point.

[0027] Among them, calculating the FPFH similarity of the source point and the target point is to calculate the similarity of each point in the source point cloud and each point in the target point cloud, and taking the maximum similarity as the score of the point. For example, there are 4 points in the source point cloud and 5 points in the target point cloud. The FPFH similarities of point 1 in the source point cloud and the 5 points in the target point cloud are 0.1, 0.2, 0.1, 0.3, and 0.5 respectively. Then the maximum similarity of point 1 is taken as the score of 0.5; similarly, the maximum similarity of the point in the target point cloud and each point in the source point cloud is taken as the score of the point in the target point cloud. Then calculate the local density of the point in the original point cloud, and use the local density to obtain the weighted value. In one embodiment, the larger the local density, the larger the weighted value. The score is weighted by the weighted value to obtain the importance score of the point.

[0028] After obtaining the importance score of the point, the source point cloud and the target point cloud are sorted according to the importance score of the point, the number of iterations is obtained, and the proportion of points retained in this iteration is determined according to the number of iterations. In one embodiment, the greater the number of iterations, the greater the proportion of retained points. For example, in the first iteration, the retained proportion is 20%, and in the second iteration, the retained proportion is 30%. According to the proportion, the point with the highest importance score is selected from the sorted source point cloud as the point of the source point cloud, and according to the proportion, the point with the highest importance score is selected from the sorted target point cloud as the point of the target point cloud; for example, if the retention ratio is 20%, the first 20% of the points are selected from the sorted source point cloud as the points of the source point cloud after downsampling, and similarly, the first 20% of the points are selected from the sorted target point cloud as the points of the target point cloud after downsampling. Obviously, the above sorting is sorted in order from large to small importance scores.

[0029] In the CPD algorithm, the matching probability is the matching situation between the source point and the target point, which is expressed by the formula Calculate the matching probability, where A is the transformation matrix, t is the translation vector, is the point of the source point cloud, is the point of the target point cloud, is the normalization factor. The source points with statistical matching probability lower than the threshold are regarded as noise points. That is, if the matching probability of a source point and all target points is less than the threshold, for example, 0.2, then this source point is regarded as a noise point.

[0030] In one embodiment, the noise weight of the next iteration is calculated by adopting a dynamic noise estimation method according to the number of noise points, specifically: Calculate the ratio of the number of noise points to the number of points in the source point cloud in this iteration; The neighborhood of the source point in the original source point cloud is obtained, and the normal difference between the source point and the neighborhood is calculated, and an adjustment coefficient is obtained according to the normal difference. The noise weight of the source point in the next iterative calculation is obtained by the adjustment coefficient and the ratio.

[0031] The overall noise level is estimated by the proportion of noise points identified in the current iteration; and an adjustment coefficient is calculated based on the normal difference between each point and its neighborhood. The adjustment coefficient is used to reflect the local geometric changes of the point. Points with larger differences will be given higher noise weights, thereby reducing the dependence on these points in the next iterative calculation.

[0032] In one embodiment, obtaining the adjustment coefficient according to the normal difference is specifically: calculating the average value of the dot product of the unit normal vectors of the source point and each point in the neighborhood; and subtracting the average value from 1 as the adjustment coefficient.

[0033] For example, in this iteration, the source point cloud has 100 points, 10 noise points are identified, and the noise ratio is 10%; for source point i, find its neighborhood in the original source point cloud and calculate the normal difference. For example, the normal difference is 0.3, and the adjustment coefficient is 1-0.3=0.7. Then, in the next iteration, the noise weight of source point i = noise ratio × adjustment coefficient = 0.1×0.7=0.07.

[0034] S3, optimizes the transformation matrix and translation vector based on the matching probability, determines the change of the transformation matrix and translation vector relative to the last calculated transformation matrix and translation vector, and stops updating the transformation matrix and translation vector when the change meets the preset conditions; uses the transformation matrix and translation vector to transform the source point cloud to complete the alignment of the source point cloud and the target point cloud.

[0035] The CPD algorithm is a process of continuously iterating and updating the transformation matrix and translation vector. After each iteration, the change of the transformation matrix and translation vector relative to the last calculated transformation matrix and translation vector is determined. When the change meets the preset conditions, the transformation matrix and translation vector are stopped from being updated. The source point cloud is transformed using the transformation matrix and translation vector to complete the registration of the source point cloud and the target point cloud. Among them, the optimization of the transformation matrix and translation vector based on the matching probability belongs to the steps in the CPD algorithm, which is a prior art and will not be repeated here.

[0036] In a first embodiment of the present invention, a laser point cloud registration system for autonomous driving is provided. Figure 2As shown, the system includes the following modules: The initialization module is used to obtain the source point cloud and target point cloud captured by the lidar in autonomous driving, initialize the noise weight based on the density of the source point cloud and the target point cloud, and calculate the fast point feature histogram of the source point cloud and the target point cloud; The downsampling module is used to downsample the source point cloud and the target point cloud according to the fast point feature histogram and the number of iterations, calculate the matching probability of the source point in the source point cloud to the target point in the target point cloud, count the source points with matching probability lower than the threshold as noise points, and calculate the noise weight of the next iteration by dynamic noise estimation according to the number of noise points; The registration module is used to optimize the transformation matrix and translation vector based on the matching probability, determine the change of the transformation matrix and translation vector relative to the last calculated transformation matrix and translation vector, and stop updating the transformation matrix and translation vector when the change meets the preset conditions; use the transformation matrix and translation vector to transform the source point cloud to complete the registration of the source point cloud and the target point cloud.

[0037] Preferably, the noise weight is initialized based on the density of the source point cloud and the target point cloud, specifically: For each point in the source point cloud and the target point cloud, calculate the density of points in its neighborhood; After calculating the local density of each point in the source point cloud and the target point cloud, the density distribution is obtained, the mean of the density is obtained, and the initialization noise weight is obtained using the mean of the density.

[0038] Preferably, downsampling the source point cloud and the target point cloud according to the fast point feature histogram and the number of iterations is specifically: Evaluate the importance score of each point based on its fast point feature histogram feature; Sort the source point cloud and the target point cloud respectively according to the importance scores of the points, obtain the number of iterations, and determine the proportion of points retained in this iteration according to the number of iterations; According to the ratio, the point with the highest importance score is selected from the sorted source point cloud as the point of the source point cloud, and according to the ratio, the point with the highest importance score is selected from the sorted target point cloud as the point of the target point cloud.

[0039] Preferably, the noise weight of the next iteration is calculated by adopting a dynamic noise estimation method according to the number of noise points, specifically: Calculate the ratio of the number of noise points to the number of points in the source point cloud in this iteration; The neighborhood of the source point in the original source point cloud is obtained, and the normal difference between the source point and the neighborhood is calculated, and an adjustment coefficient is obtained according to the normal difference. The noise weight of the source point in the next iterative calculation is obtained by the adjustment coefficient and the ratio.

[0040] Preferably, the adjustment coefficient is obtained according to the normal difference, specifically: Calculate the average of the dot products of the unit normal vectors of the source point and each point in the neighborhood; The adjustment factor is 1 minus the average value.

[0041] Preferably, the importance score of the point is evaluated based on the fast point feature histogram feature of each point, specifically: Calculate the similarity of the fast point feature histograms of the source point and the target point. For each point in the source point cloud and the target point cloud, take the maximum similarity as the score of the point. The local density of the point in the original source point cloud or the original target point cloud is obtained, and the weighted value is obtained by using the local density. The score is weighted by using the weighted value to obtain the importance score of the point.

[0042] Through the description of the above implementation methods, technicians in this field can clearly understand that each implementation method can be implemented by adding a necessary general hardware platform, and of course can also be implemented by combining hardware and software. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a computer product, and the present invention can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it, and other embodiments may also be used. Although the present invention has been described in detail with reference to the aforementioned embodiments, ordinary technicians in this field should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A laser point cloud registration method for autonomous driving, characterized in that: The method comprises the following steps: Obtain the source point cloud and target point cloud taken by the LiDAR in autonomous driving, initialize the noise weight based on the density of the source point cloud and target point cloud, and calculate the fast point feature histogram of the source point cloud and target point cloud; The source point cloud and the target point cloud are downsampled according to the fast point feature histogram and the number of iterations, the matching probability of the source point in the source point cloud to the target point in the target point cloud is calculated, the source points with matching probability lower than the threshold are counted as noise points, and the noise weight of the next iteration is calculated by dynamic noise estimation according to the number of noise points; Based on the matching probability, the transformation matrix and translation vector are optimized, and the change of the transformation matrix and translation vector relative to the last calculated transformation matrix and translation vector is determined. When the change meets the preset conditions, the transformation matrix and translation vector are stopped from being updated. The source point cloud is transformed using the transformation matrix and translation vector to complete the registration of the source point cloud and the target point cloud.

2. The method according to claim 1, characterized in that The noise weight is initialized based on the density of the source point cloud and the target point cloud, specifically: For each point in the source point cloud and the target point cloud, calculate the density of points in its neighborhood; After calculating the local density of each point in the source point cloud and the target point cloud, the density distribution is obtained, the mean of the density is obtained, and the initialization noise weight is obtained using the mean of the density.

3. The method according to claim 1, characterized in that The downsampling of the source point cloud and the target point cloud according to the fast point feature histogram and the number of iterations is specifically as follows: Evaluate the importance score of each point based on its fast point feature histogram feature; Sort the source point cloud and the target point cloud respectively according to the importance scores of the points, obtain the number of iterations, and determine the proportion of points retained in this iteration according to the number of iterations; According to the ratio, the point with the highest importance score is selected from the sorted source point cloud as the point of the source point cloud, and according to the ratio, the point with the highest importance score is selected from the sorted target point cloud as the point of the target point cloud.

4. The method according to claim 1, characterized in that The noise weight of the next iteration is calculated by adopting a dynamic noise estimation method according to the number of noise points, specifically: Calculate the ratio of the number of noise points to the number of points in the source point cloud in this iteration; The neighborhood of the source point in the original source point cloud is obtained, and the normal difference between the source point and the neighborhood is calculated, and an adjustment coefficient is obtained according to the normal difference. The noise weight of the source point in the next iterative calculation is obtained by the adjustment coefficient and the ratio.

5. The method according to claim 4, characterized in that The adjustment coefficient is obtained according to the normal difference, specifically: Calculate the average of the dot products of the unit normal vectors of the source point and each point in the neighborhood; The adjustment factor is 1 minus the average value.

6. The method according to claim 3, characterized in that The importance score of the point is evaluated based on the fast point feature histogram feature of each point, specifically: Calculate the similarity of the fast point feature histograms of the source point and the target point. For each point in the source point cloud and the target point cloud, take the maximum similarity as the score of the point. The local density of the point in the original source point cloud or the original target point cloud is obtained, and the weighted value is obtained by using the local density. The score is weighted by using the weighted value to obtain the importance score of the point.

7. A laser point cloud registration system for autonomous driving, characterized in that: The system includes the following modules: The initialization module is used to obtain the source point cloud and target point cloud captured by the lidar in autonomous driving, initialize the noise weight based on the density of the source point cloud and the target point cloud, and calculate the fast point feature histogram of the source point cloud and the target point cloud; The downsampling module is used to downsample the source point cloud and the target point cloud according to the fast point feature histogram and the number of iterations, calculate the matching probability of the source point in the source point cloud to the target point in the target point cloud, count the source points with matching probability lower than the threshold as noise points, and calculate the noise weight of the next iteration by dynamic noise estimation according to the number of noise points; The registration module is used to optimize the transformation matrix and translation vector based on the matching probability, determine the change of the transformation matrix and translation vector relative to the last calculated transformation matrix and translation vector, and stop updating the transformation matrix and translation vector when the change meets the preset conditions; use the transformation matrix and translation vector to transform the source point cloud to complete the registration of the source point cloud and the target point cloud.

8. The system according to claim 7, characterized in that The noise weight is initialized based on the density of the source point cloud and the target point cloud, specifically: For each point in the source point cloud and the target point cloud, calculate the density of points in its neighborhood; After calculating the local density of each point in the source point cloud and the target point cloud, the density distribution is obtained, the mean of the density is obtained, and the initialization noise weight is obtained using the mean of the density.

9. The system according to claim 7, characterized in that The downsampling of the source point cloud and the target point cloud according to the fast point feature histogram and the number of iterations is specifically as follows: According to the distribution of FPFH features of each point, the importance score of the evaluation point is evaluated; Sort the source point cloud and the target point cloud respectively according to the importance scores of the points, obtain the number of iterations, and determine the proportion of points retained in this iteration according to the number of iterations; According to the ratio, the point with the highest importance score is selected from the sorted source point cloud as the point of the source point cloud, and according to the ratio, the point with the highest importance score is selected from the sorted target point cloud as the point of the target point cloud.

10. The system according to claim 7, characterized in that The noise weight of the next iteration is calculated by adopting a dynamic noise estimation method according to the number of noise points, specifically: Calculate the ratio of the number of noise points to the number of points in the source point cloud in this iteration; The neighborhood of the source point in the original source point cloud is obtained, and the normal difference between the source point and the neighborhood is calculated, and an adjustment coefficient is obtained according to the normal difference. The noise weight of the source point in the next iterative calculation is obtained by the adjustment coefficient and the ratio.

Images