Registration method based on super voxel segmentation and bidirectional nearest neighbor distance ratio matching
By using supervoxel segmentation and bidirectional nearest neighbor distance ratio matching, the problem of insufficient accuracy and speed in point cloud registration in existing technologies is solved, achieving high-precision and fast point cloud registration, especially with strong robustness in noisy environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-17
- Publication Date
- 2026-03-03
AI Technical Summary
Existing 3D point cloud registration methods cannot meet the higher requirements for registration accuracy and speed of target point clouds acquired by sensors such as lidar.
A registration method based on supervoxel segmentation and bidirectional nearest neighbor ratio matching is adopted, including supervoxel segmentation algorithm to extract target feature points of stable structures, non-iterative thresholding denoising, FPFH feature description and bidirectional nearest neighbor ratio initial registration, and point cloud accurate registration based on dual-level thresholds.
It significantly improves the accuracy and speed of point cloud registration, especially showing strong robustness in noisy environments, significantly reducing registration error and improving registration accuracy by 68.44%~67.73%.
Smart Images
Figure CN115205352B_ABST
Abstract
Description
Technical fields:
[0001] This invention belongs to the field of radar signal processing technology and relates to a registration method based on supervoxel segmentation and bidirectional nearest neighbor range ratio matching. Background technology:
[0002] Due to limitations in sensor penetration and target occlusion, the sensor receiving system acquires localized target information from different imaging perspectives. To facilitate target identification and subsequent effective processing of point cloud data, coordinate transformation is required to quickly and accurately register consecutive frames, thereby obtaining richer and more complete target information.
[0003] Point cloud registration typically involves two stages: initial registration and precise registration. Initial registration involves preprocessing the point cloud data from different perspectives to the same coordinate system after point cloud denoising and filtering, sacrificing some registration accuracy. Common initial registration methods include point-based, line-based, and area-based methods. Precise registration involves refining the initial transformation relationship to further reduce registration errors and meet predetermined registration requirements, achieving accurate image matching. Common fine registration methods include iterative approximation, random sample consistency, and normal distribution methods.
[0004] There are many methods for 3D point cloud registration. Because some sensors, such as lidar and millimeter-wave radar, can obtain not only distance information along the target imaging direction but also spatial distribution information perpendicular to the imaging direction, but are limited by the imaging constraints of the sensor's ranging blind zone and the real-time requirements of the carrier platform, higher demands are placed on the registration accuracy and speed of target point clouds acquired by lidar and other sensors. Existing 3D point cloud registration methods cannot meet the higher requirements for registration accuracy and speed for target point clouds acquired by lidar and other sensors. Summary of the Invention:
[0005] The purpose of this invention is to provide a registration method based on supervoxel segmentation and bidirectional nearest neighbor distance ratio matching, which overcomes the limitations of existing 3D point cloud registration methods in meeting the higher requirements for registration accuracy and speed of target point clouds acquired by sensors such as LiDAR. This invention utilizes supervoxel segmentation to accurately extract target feature points with stable structures, eliminates interference from drift noise voxels, and improves registration accuracy using bidirectional nearest neighbor distance ratio matching.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] A registration method based on supervoxel segmentation and bidirectional nearest neighbor distance ratio matching, characterized by the following steps:
[0008] Step 1: Extract target feature points with stable structures using the hypervoxel segmentation algorithm;
[0009] Step 2: While performing supervoxel segmentation, non-iterative threshold denoising is performed using point cloud thickness layering;
[0010] Step 3: Feature description is performed using FPFH, and initial registration of the point cloud is performed using the bidirectional nearest neighbor distance ratio method;
[0011] Step 4: Perform point cloud registration using a point cloud registration method based on a two-level threshold.
[0012] Step two includes the following steps:
[0013] Step 1: Normalize the meshed point cloud VC1 using the maximum value to obtain the point cloud VC2;
[0014] Step 2: Sort point cloud VC2 in ascending order according to the target imaging distance information to obtain point cloud VC3; rearrange the point cloud in ascending order according to the target Z-axis direction to facilitate subsequent layering of the point cloud according to thickness.
[0015] Step 3: Divide the point cloud VC3 into layers according to thickness, and calculate the layer thickness as shown in equation (1):
[0016]
[0017] In the formula, th represents the thickness of each point cloud layer, and S... xoy ,S yoz ,S xoz N represents the projected area of the smallest bounding box on xoy, yoz, and xoz. v2 This represents the total number of VC3 voxels in the point cloud;
[0018] Step 4: Calculate the number of layers in the point cloud VC3, as shown in equation (2):
[0019]
[0020] In the formula, H represents the number of layers in the point cloud, and L... P This represents the length of the target distance information in the point cloud along the Z-axis direction.
[0021] Step 5: After the point cloud is layered using VC3, extract the voxels of each layer and calculate the minimum bounding box of each voxel based on a threshold. The threshold expression is shown in equation (3):
[0022]
[0023] In the formula, T represents the threshold and γ represents the threshold coefficient;
[0024] Step 6: Let V be the j-th voxel in the i-th layer (i∈[1,H]). i,j (x i,j ,y i,j ,z i,j ), then V i,j The vertex coordinates of the minimum bounding box are in the range [(x i,j -T,x i,j +T,),(y i,j -T,y i,j +T),(z i,j -T,z i,j +T)];If the number of voxels in the minimum bounding box of a voxel exceeds 4, it is retained; otherwise, it is considered a dense drift noise voxel and is removed; after denoising, the point cloud VC4 is obtained.
[0025] Step 7: Restore VC4 to the original point cloud size to obtain point cloud VC5, which is the selected correct seed voxel.
[0026] Step three includes the following steps:
[0027] Step 1: Large convergence threshold
[0028] Considering the direct impact of ranging accuracy on registration error, the large convergence threshold model is constructed as follows:
[0029]
[0030] In the formula, Vr represents voxel resolution; RA represents ranging accuracy.
[0031] Step 2: Small convergence threshold
[0032] The overlap between two point clouds to be registered is expressed as follows:
[0033]
[0034] In the formula, N c N represents the number of overlapping points in a point cloud. t This represents the total number of points in the point cloud.
[0035] Step 3: The larger the CD value of two point clouds, the smaller the distance between corresponding point pairs, and the distance between point pairs in the imaging direction and perpendicular to the imaging direction decreases accordingly; the constructed small threshold convergence model is as follows:
[0036]
[0037] When the registration error is less than CTH min Registration ends when the registration error exceeds CTH. minAnd smaller than CTH max At that time, remove erroneous pairs that do not meet the conditions;
[0038] Step 4: Set a distance threshold to determine if the corresponding point pairs are incorrect. Construct the distance threshold model as follows:
[0039]
[0040] When the nearest distance between corresponding point pairs is greater than DTH, these corresponding point pairs are removed as incorrect matching point pairs; when it is less than DTH, corresponding points are searched again, and the overlap and registration error are updated.
[0041] Compared with the prior art, the advantages and effects of the present invention are as follows:
[0042] 1. This invention is based on supervoxel segmentation and employs a registration method combining initial registration and precise registration. First, the 3D point cloud data acquired by the sensor is preprocessed by voxelizing and meshing the point cloud. Then, non-iterative thresholding and k-means clustering are used to generate supervoxels with stable structures, achieving supervoxel clustering segmentation. This effectively reduces the processing scale of the point cloud data and the complexity of subsequent registration processing, while also effectively extracting the stable structure of the target. Next, supervoxel features are extracted from the normal vectors, and the point cloud is characterized using the FPFH feature descriptor. Initial matching is then performed using the bidirectional nearest neighbor distance ratio. Finally, an improved adaptive dual-threshold algorithm is used to complete the precise registration of the point cloud. This registration method provides a foundation for subsequent point cloud target identification, tracking, localization, and attitude estimation.
[0043] 2. This invention uses supervoxel clustering to divide the point cloud into layers according to thickness, and proposes a non-iterative threshold denoising method to effectively remove drift noise voxels.
[0044] 3. A bidirectional nearest neighbor distance ratio-based initial registration method based on FPFH feature extraction is proposed, which effectively improves the number of correct matching point pairs. When the point cloud is not segmented into supervoxels, the registration error of this initial registration method is 2.4472 × 10⁻⁶. -4 The accuracy is significantly higher than that of one-way matching.
[0045] 4. A point cloud accurate registration method based on dual-level thresholding is proposed. When the point cloud has not undergone super-voxel segmentation, the registration error of the method of FPFH + bidirectional nearest neighbor distance ratio initial registration + dual-level thresholding is 3.6786 × 10⁻⁶. -34 The accuracy is significantly higher than that of the FPFH + bidirectional nearest neighbor distance ratio initial matching + ICP method.
[0046] 5. The point cloud accurate registration method proposed in this invention, which combines super-voxel segmentation, bidirectional nearest neighbor distance ratio initial registration, and dual-level thresholding, achieves registration errors of 3.8025 × 10⁻⁶ when the noise percentage is 6% and 10%, respectively. -5 and 3.5942×10 -5 The registration accuracy was improved by 68.44% and 67.73% respectively, and the noise robustness was strong. Attached image description:
[0047] Figure 1 The results are experimental findings using the horse model in the noise ratio range of [0 0.2].
[0048] Figure 2 The results are experimental findings obtained by using the horse model to conduct experiments in the voxel resolution range of [0.001 0.0025].
[0049] Figure 3 The results are the noise-free horse model data and the visualization results with 6% noise added.
[0050] Among them, (a) the original image without noise; (b) the original image with 6% noise.
[0051] Figure 4 The registration results of various algorithms for the horse model under noise-free conditions are shown.
[0052] Among them, (a) ICP; (b) FPFH+ICP; (c) FPFH+dual threshold; (d) the algorithm of this invention.
[0053] Figure 5 The registration results of the horse model using various algorithms after adding 6% noise are shown.
[0054] Among them, (a) ICP; (b) FPFH+ICP; (c) FPFH+dual threshold; (d) the algorithm of this invention. Detailed implementation method:
[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0056] This invention discloses a registration method based on supervoxel segmentation and bidirectional nearest neighbor ratio matching. The method accurately extracts target feature points with stable structures using supervoxel segmentation, eliminates interference from drift noise voxels, and improves registration accuracy using bidirectional nearest neighbor ratio matching. The invention includes the following steps: First, it extracts target feature points with stable structures using a supervoxel segmentation algorithm, and simultaneously performs non-iterative threshold denoising using point cloud thickness layering; then, it uses FPFH for feature description and performs initial registration of the point cloud using the bidirectional nearest neighbor ratio method; finally, it achieves accurate point cloud registration using a two-level threshold-based method.
[0057] The present invention specifically includes the following steps:
[0058] Step 1: Extract target feature points with stable structures using the hypervoxel segmentation algorithm;
[0059] Step 2: While performing supervoxel segmentation, non-iterative threshold denoising is performed using point cloud thickness layering;
[0060] Step 3: Feature description is performed using FPFH, and initial registration of the point cloud is performed using the bidirectional nearest neighbor distance ratio method;
[0061] Step 4: Point cloud registration was achieved using a point cloud registration method based on a two-level threshold.
[0062] The specific algorithm for step two above is as follows:
[0063] Step 1: Normalize the meshed point cloud VC1 using the maximum value to obtain point cloud VC2.
[0064] Step 2: Arrange point cloud VC2 in ascending order according to the target imaging distance information direction to obtain point cloud VC3. This invention rearranges the point cloud in ascending order according to the target Z-axis direction, which facilitates subsequent layering of the point cloud according to thickness.
[0065] Step 3: Divide the point cloud VC3 into layers according to thickness, and calculate the layer thickness as shown in equation (1):
[0066]
[0067] In the formula, th represents the thickness of each point cloud layer, and S... xoy ,S yoz ,S xoz N represents the projected area of the smallest bounding box on xoy, yoz, and xoz. v2 This represents the total number of VC3 voxels in the point cloud.
[0068] Step 4: Calculate the number of layers in the point cloud VC3, as shown in equation (2):
[0069]
[0070] In the formula, H represents the number of layers in the point cloud, and L... P This represents the length of the target distance information in the point cloud, which in this invention is the length of the target in the Z-axis direction.
[0071] Step 5: After the point cloud is layered using VC3, extract the voxels of each layer and calculate the minimum bounding box of each voxel based on a threshold. The threshold expression is shown in equation (3):
[0072]
[0073] In the formula, T represents the threshold and γ represents the threshold coefficient.
[0074] Step 6: Let V be the j-th voxel in the i-th layer (i∈[1,H]). i,j (x i,j ,y i,j ,z i,j ), then V i,j The vertex coordinates of the minimum bounding box are in the range [(x i,j -T,x i,j +T,),(y i,j -T,y i,j +T),(z i,j -T,z i,j If the number of voxels within the minimum bounding box of a voxel exceeds four, it is retained; otherwise, it is considered a dense drift noise voxel and is removed. The denoised result is the point cloud VC4.
[0075] Step 7: Restore VC4 to its original point cloud size to obtain point cloud VC5. VC5 is the selected correct seed voxel.
[0076] The specific steps of step three above are as follows:
[0077] Step 1: Large convergence threshold
[0078] Considering the direct impact of ranging accuracy on registration error, the large convergence threshold model is constructed as follows:
[0079]
[0080] In the formula, RA represents the ranging accuracy.
[0081] Step 2: Small convergence threshold
[0082] The overlap between two point clouds to be registered is expressed as follows:
[0083]
[0084] In the formula, N c N represents the number of overlapping points in a point cloud. t This represents the total number of points in the point cloud.
[0085] Step 3: The larger the CD value of two point clouds, the smaller the distance between corresponding point pairs, and the distance between point pairs along the imaging direction and perpendicular to the imaging direction decreases accordingly. Therefore, the constructed small threshold convergence model is as follows:
[0086]
[0087] When the registration error is less than CTH min Registration ends when the registration error exceeds CTH. min And smaller than CTH max When doing so, error pairs that do not meet the conditions should be removed.
[0088] Step 4: Set a distance threshold to determine if the corresponding point pairs are incorrect. Construct the distance threshold model as follows:
[0089]
[0090] When the nearest distance between corresponding point pairs is greater than DTH, these corresponding point pairs are removed as incorrect matching point pairs; when it is less than DTH, corresponding points are searched again, and the overlap and registration error are updated.
[0091] Example:
[0092] Example 1: Bidirectional matching using a horse model. In horse model matching, the target point cloud is obtained by rotating the original point cloud by 60 degrees in space. The simulation parameter settings are shown in Table 1.
[0093] Table 1 Registration Parameters
[0094]
[0095] Specifically, the following steps are included:
[0096] Step 1: Add Gaussian noise
[0097] Due to the influence of factors such as the measurement environment, sensor measurement errors, and surface characteristics of the object being measured, the acquired point cloud data inevitably contains noise. In a typical model library, a horse model was selected and Gaussian noise was added to evaluate the robustness of the registration algorithm proposed in this invention. The percentage of correctly matched point pairs was used as the evaluation metric. The experiment was repeated 10 times, and the average result was taken.
[0098] In this invention, "noise ratio" refers to the proportion of added Gaussian noise points to the total number of points in the model point cloud. The total number of points in the point cloud is the sum of the model point cloud points and the Gaussian noise points. Table 2 shows that, since FPFH primarily relies on the normal vector information of the point cloud for initial matching during the bidirectional nearest neighbor distance ratio (BNFH) process, the number of relevant point pairs is not high for horse data models with low geometric features. When the noise ratio is between [0, 0, 2], the number of correctly matched point pairs for the horse model is between [4, 3, 5, 6], indicating that the initial registration method of BNFH has relatively stable robustness. The average distance between corresponding points in the point cloud data is usually used as the point cloud resolution. This invention changes the voxel resolution R... v Uniform sampling is performed on the point cloud, while adhering to a grid resolution of R. m =8×R v This reduces the point cloud resolution. As the voxel resolution increases, the number of seed voxels for meshing decreases, and consequently, the number of relevant point pairs decreases. When the voxel resolution varies in the range of [0.001-0.0025], the number of correctly matched point pairs in the Maas model varies between [8-43], indicating that the algorithm has strong resolution at different imaging resolutions.
[0099] Experiments were conducted using the horse model in the noise ratio range of [0 0.2] and the voxel resolution range of [0.001 0.0025] to further verify the robustness performance. The experimental results are as follows: Figure 1 and Figure 2 As shown.
[0100] Table 2 Comparison of point cloud resolution of typical models
[0101]
[0102] Example 2: Comparative Analysis of Registration Time
[0103] The simulation was performed using MATLAB R2018b on a computer system running Windows 10 64-bit operating system, with an Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz and 3.19GHz processors and 8GB of memory. Taking the horse model as an example, the point cloud size was 48485 points. The following algorithms were run 50 times each, and the algorithm running times are shown in Table 3.
[0104] As can be seen from Table 3, when the point cloud data is initially matched using FPFH+ bidirectional nearest neighbor distance ratio, the fine matching uses ICP and dual-level threshold algorithms respectively. The latter reduces the registration time by 0.7984 seconds and improves the registration accuracy by 74.2%. When the point cloud is initially matched using FPFH+ unidirectional nearest neighbor distance and FPFH+ bidirectional nearest neighbor distance ratio respectively, the latter has a longer registration time, but the error is significantly reduced and the registration accuracy is improved by 99.4%.
[0105] Table 3 Comparison of Registration Time
[0106]
[0107] Example 3: Comparative Analysis of Registration Accuracy
[0108] Scan data such as Figure 3 As shown. Under noise-free conditions, the registration results of various algorithms are as follows. Figure 4 As shown. When the noise percentage is 6%, the registration results of various algorithms are as follows. Figure 5 As shown. From Figure 5 As can be seen, compared with other methods, the method proposed in this invention has better noise robustness and can complete the matching better. Compared with other methods, the method proposed in this invention has the smallest registration error when the noise ratio is 6% and 10%, and the registration accuracy is improved by 68.44% and 67.73% respectively compared with the traditional ICP algorithm.
[0109] Table 4 Comparison of Registration Accuracy under Noise Conditions
[0110]
[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. All equivalent structural changes made based on the description and drawings of the present invention should be included within the scope of patent protection of the invention.
Claims
1. A registration method based on super voxel segmentation and bidirectional nearest neighbor distance ratio matching, characterized in that: Comprise the following steps: Step one: using super voxel segmentation algorithm to extract the target feature point with stable structure; Step two: in the super voxel segmentation at the same time using point cloud thickness layer non-iterative threshold denoising; Step three: using FPFH for feature description, using bidirectional nearest neighbor distance ratio method to point cloud has carried out the initial registration; Step four: using the point cloud accurate registration method based on double threshold realizes point cloud registration; Step four comprises the following steps: Step 1: large convergence threshold Considering the direct influence degree of ranging accuracy on registration error, the large convergence threshold model is constructed as follows: In the formula, Vr represents the voxel resolution; RA represents the ranging accuracy; Step 2: small convergence threshold The coincidence degree of two pieces of point cloud to be registered is represented as follows: In the formula, N c represents the number of coincident points in the point cloud, N t represents the total number of points in the point cloud; Step 3: when the CD value of two pieces of point cloud is larger, the distance between corresponding point pairs is smaller, and the distance between imaging direction and perpendicular to imaging direction is also reduced; The small threshold convergence model is constructed as follows: When the registration error is less than CTH min , the registration ends; when the registration error is greater than CTH min and less than CTH max , the error point pairs that do not meet the condition are rejected. Step 4: set the distance threshold to judge whether it is an incorrect corresponding point pair, and the distance threshold model is constructed as follows: When the nearest distance between corresponding point pairs is greater than DTH, these corresponding point pairs are removed as incorrect matching point pairs; When less than DTH, find the corresponding points again, and update the coincidence degree and registration error.
2. The registration method based on super voxel segmentation and bidirectional nearest neighbor distance ratio matching according to claim 1, characterized in that: Step two comprises the following steps: Step 1: the grid point cloud VC1 is normalized by using the maximum value, and the point cloud VC2 is obtained; Step 2: the point cloud VC2 is arranged in ascending order according to the target imaging distance information direction, and the point cloud VC3 is obtained; The point cloud is arranged in ascending order according to the target Z axis direction, which is convenient for subsequent layering processing of point cloud according to thickness; Step 3: the point cloud VC3 is layered according to thickness, and the layering thickness is calculated as shown in formula (1): In the formula, th represents the thickness of each layer of point cloud, S xoy ,S yoz ,S xoz represents the projection area of the minimum bounding box on xoy, yoz and xoz, N v2 represents the total number of point cloud VC3 voxels; Step 4: calculate the number of layers of point cloud VC3, as shown in formula (2): In the formula, H represents the number of layers of the point cloud, L P represents the length of the target distance information direction in the point cloud, and is the length of the target Z-axis direction. Step 5: after the point cloud VC3 is layered, each layer of voxel is taken out, and the minimum bounding box of each layer of voxel is calculated according to the threshold value, and the threshold value expression is shown as formula (3): In the formula, T represents the threshold value, and γ represents the threshold value coefficient; Step 6: Let the jth voxel in the ith (i∈[1,H]) layer be V i,j (x i,j ,y i,j ,z i,j ), then the vertex coordinate range of the minimum bounding box of V i,j is [(x i,j -T,x i,j +T),(y i,j -T,y i,j +T),(z i,j -T,z i,j +T)]; if the number of voxels in the minimum bounding box of the voxel exceeds 4 or more voxels, it is retained, otherwise it is considered to be a tight drift noise voxel, which is removed; after denoising, a point cloud VC4 is obtained; Step 7: restore VC4 to the original point cloud size, and get point cloud VC5, VC5 is the correct seed voxel selected.