A binary descriptor acquisition method based on local voxel coding

Abstract

The application discloses a binary descriptor acquisition method based on local voxel coding, comprising the following steps: firstly, extracting local point clouds around key points, calculating a local reference frame from the local point clouds, and converting the local point clouds into the local reference frame; then, voxelizing the converted local point clouds, comparing the relationship between occupied voxels for each layer, obtaining a binary string, and concatenating all the binary strings to obtain a binary descriptor. After extracting key points of two point clouds to be registered, the descriptor proposed in the application is calculated on the key points. Through the descriptor, the pairing between the two point clouds can be established. The better the matching effect of the descriptor is, the more correct pairings can be established, and the point cloud registration is ensured to be successful.

Description

Technical Field

[0001] This invention relates to the field of 3D point cloud data processing technology, specifically, a method for obtaining binary descriptors based on local voxel encoding. Background Technology

[0002] With the development of 3D acquisition technology, point cloud data has become increasingly common, serving as a primary data source for many practical applications, such as scene or model reconstruction, 3D target recognition, and indoor / outdoor location identification. Point cloud registration is a fundamental problem that needs to be solved in these applications, and therefore has received widespread attention, becoming a hot research topic. Local shape descriptors (LSDs) have been widely used in point cloud registration, and LSD-based point cloud registration methods have been extensively studied.

[0003] Currently, numerous descriptors have been proposed, which can be categorized into floating-point descriptors and binary descriptors. Floating-point descriptors mainly include spin images (SI), divisional local feature statistics (DLFS), histogram of distances (HoD), signatures of histograms of orientations (SHOT), spatial and geometric histogram (SGH), kernel density descriptor (KDD), and weighted height image (WHI). Among these, SI, DLFS, and HoD are descriptors without local surface-rendering (LRF). These descriptors only encode the geometric information of the local surface, discarding spatial information, and therefore generally have poor descriptive performance. SHOT, SGH, KDD, and WHI are descriptors with LRF. Because LRF provides a way to encode spatial information, these descriptors consider both geometric and spatial information. However, these descriptors still cannot meet the needs of practical applications and require further improvement in performance. Furthermore, floating-point descriptors generally require more memory, and descriptor matching takes longer.

[0004] Binary descriptors mainly include binary SHOT (B-SHOT), binary HoD (B-HoD), binary shape context (BSC), local voxelized structure (LoVS), and rotational silhouettemaps (RSM). B-SHOT, B-HoD, and BSC are obtained by binarying their corresponding floating-point descriptors. During binarying, some information is inevitably lost, resulting in these descriptors having poorer descriptive power than their floating-point counterparts. LoVS and RSM are natural binary descriptors, not obtained through binarying, and achieve excellent descriptive power. However, RSM projects the local point cloud onto the three planes of the LRF, leading to some information loss. LoVS directly encodes local surface information in 3D space, achieving better descriptive power than RSM, but it simply encodes local voxels without considering the relationships between voxels. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a binary descriptor based on local voxel encoding. This descriptor is a natural binary descriptor that encodes local surface information in 3D space, takes into account the relationships between voxels, and has better descriptiveness.

[0006] To achieve the above objectives, the present invention provides the following technical solution.

[0007] A method for obtaining binary descriptors based on local voxel encoding includes the following steps:

[0008] Step S1: Calculate the local reference frame based on the local point cloud around the key point, and convert the local point cloud into the local reference frame;

[0009] Step S2: Convert the transformed local point cloud into voxels, calculate the binary string, and then concatenate all the binary strings to obtain the binary descriptor.

[0010] Step S1 involves calculating a local reference frame based on the local point cloud surrounding the key points and then converting the local point cloud into the local reference frame. The specific process is as follows:

[0011] Extract the local point cloud around the keypoint, with the keypoint as the center and a radius of... Extracting local point clouds from circles, Set it to 15pr, where pr represents the point cloud resolution;

[0012] Firstly, the key point distance is less than... Calculate the normal vectors of the points within the matrix and construct the covariance matrix.

[0013] (1);

[0014] In the above formula, This indicates a distance to a key point that is less than point, This represents the average value of these points;

[0015] Performing eigenvalue decomposition on the covariance matrix yields three eigenvalues ​​and three corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue is the normal vector. ;

[0016] Based on the normal vector The z-axis of the local reference frame (LRF) is determined using the following formula.

[0017] (2);

[0018] In the above formula, Indicate key points, This represents a point in a local point cloud.

[0019] The x-axis of the local reference frame (LRF) is calculated as follows:

[0020] (3);

[0021] In the above formula, It is the projection vector. and This indicates two rights; rights To improve the robustness of LRF to partial overlap, weights Improve the repeatability of LRF in flat regions;

[0022] Finally, the y-axis of the local reference frame (LRF) is calculated using the cross product.

[0023] (4);

[0024] The local point cloud is converted into the local reference frame (LRF) according to formulas (2), (3), and (4).

[0025] Step S2 involves voxelizing the converted local point cloud, calculating the binary string, and then concatenating all the binary strings to obtain a binary descriptor, the specific content of which is as follows:

[0026] Step S21: Convert the transformed local point cloud into voxels and divide it into 15×15×15 voxels of the same size. Let m be the occupied voxel and the other 15×15×15-m voxels be denoted as empty voxels.

[0027] Step S22: From top to bottom, for each layer after local point cloud voxelization, calculate the eigenvalues ​​corresponding to all symmetric voxels;

[0028] Step S23: The eigenvalues ​​obtained from calculating all symmetric voxels are combined into a binary string;

[0029] Step S24: Ignore the voxels on the main diagonal, concatenate the binary strings of all layers to obtain the binary descriptor.

[0030] In step S22, for each layer after local point cloud voxelization, the eigenvalues ​​corresponding to all symmetrical voxels are calculated. The calculation process is as follows: for two voxels that are symmetrical about the main diagonal, if they are both occupied by points or are both empty voxels, the eigenvalue is 0; if one voxel is occupied and the other is empty, the eigenvalue is 1.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] 1. The descriptor obtained by the method of the present invention directly encodes local surface information in 3D space, which can reduce information loss. At the same time, when encoding voxels, the relationship between voxels is considered, so that the descriptor contains more local surface information, and the obtained descriptor has better descriptiveness.

[0033] 2. Experimental analysis shows that the descriptors obtained by the method of the present invention are superior to traditional LoVS, TOLDI, RoPS, USC and SI descriptors, achieving better recall and precision, and better descriptor matching effect. Using the descriptors of the present invention can establish more correct matches. Attached Figure Description

[0034] Figure 1 This is a flowchart of a binary descriptor acquisition method based on local voxel encoding according to the present invention;

[0035] Figure 2 This is a schematic diagram of point cloud voxelization;

[0036] Figure 3 The RP curves for each descriptor;

[0037] Figure 4 This refers to the registration process for point clouds. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1

[0040] like Figure 1 As shown, this embodiment provides a method for obtaining binary descriptors based on local voxel encoding, including the following steps:

[0041] Step S1: Calculate the local reference frame based on the local point cloud around the key point, and convert the local point cloud into the local reference frame;

[0042] Step S2: Convert the transformed local point cloud into voxels, calculate the binary string, and then concatenate all the binary strings to obtain the binary descriptor.

[0043] Step S1 involves calculating a local reference frame based on the local point cloud surrounding the key points and then converting the local point cloud into the local reference frame. The specific process is as follows:

[0044] Extract the local point cloud around the keypoint, with the keypoint as the center and a radius of... Extracting local point clouds from circles, Set it to 15pr, where pr represents the point cloud resolution;

[0045] Firstly, the key point distance is less than... Calculate the normal vectors of the points within the matrix and construct the covariance matrix.

[0046] (1);

[0047] In the above formula, This indicates a distance to a key point that is less than point, This represents the average value of these points;

[0048] Performing eigenvalue decomposition on the covariance matrix yields three eigenvalues ​​and three corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue is the normal vector. ;

[0049] Based on the normal vector The z-axis of the local reference frame (LRF) is determined using the following formula.

[0050] (2);

[0051] In the above formula, Indicate key points, It represents a point in the local point cloud; Equation (2) is to eliminate the ambiguity of the z-axis and make the obtained LRF unique.

[0052] The x-axis of the local reference frame (LRF) is calculated as follows:

[0053] (3);

[0054] In the above formula, It is the projection vector. and This indicates two rights; rights To improve the robustness of LRF to partial overlap, weights Improve the repeatability of LRF in flat regions;

[0055] Finally, the y-axis of the local reference frame (LRF) is calculated using the cross product.

[0056] (4);

[0057] The local point cloud is converted into the local reference frame (LRF) according to formulas (2), (3), and (4).

[0058] Step S2 involves voxelizing the converted local point cloud, calculating the binary string, and then concatenating all the binary strings to obtain a binary descriptor, the specific content of which is as follows:

[0059] Step S21: Voxelize the converted local point cloud, such as... Figure 2 As shown, it is divided into 15×15×15 voxels of the same size. The red ones represent voxels occupied by points, and the others represent empty voxels.

[0060] Step S22: From top to bottom, for each layer after local point cloud voxelization, calculate the eigenvalues ​​corresponding to all symmetric voxels;

[0061] Step S23: The eigenvalues ​​obtained from the calculation of all symmetric voxels are used to form a binary string (e.g., 0011100111100101).

[0062] Step S24: Ignore the voxels on the main diagonal, concatenate the binary strings of all layers to obtain the binary descriptor.

[0063] In step S22, for each layer after local point cloud voxelization, the eigenvalues ​​corresponding to all symmetrical voxels are calculated. The calculation process is as follows: for two voxels that are symmetrical about the main diagonal, if they are both occupied by points or are both empty voxels, the eigenvalue is 0; if one voxel is occupied and the other is empty, the eigenvalue is 1.

[0064] Example 2

[0065] The following specific experiment compares the descriptor obtained by the method of the present invention with the traditional descriptor to verify the superiority of the descriptor obtained by the present invention.

[0066] The experimental process in this embodiment is as follows: The experiment uses three datasets: Space Time, UWA3M, and UWAOR. The Space Time dataset contains 8 model point clouds and 15 scene point clouds; UWAOR contains 5 model point clouds and 50 scene point clouds, and each model point cloud is registered to the scene point cloud; the UWA3M dataset contains 496 pairs of point clouds.

[0067] For each pair of point clouds in the three datasets, consistent sampling is used to extract key points. First, 1000 key points are extracted from the original point cloud. Then, the key points are transformed to the target point cloud according to the real transformation. The nearest point of these transformed key points is found in the target point cloud and used as the key point of the target point cloud to obtain the real pair (the real transformation of the three datasets is provided by the data publisher).

[0068] Descriptors for these keypoints are calculated, and the similarity of descriptors is compared using the NNSR method to establish pairs. The threshold for the NNSR method is varied from 0.1 to 1, with an interval of 0.1. Precision and recall are calculated at each threshold using the following formulas.

[0069] (5);

[0070] (6);

[0071] Plot the RP curve. If the descriptor has both high precision and good recall, the curve should fall in the upper left corner of the graph. Compare the descriptors obtained according to the method of this invention with LoVS, TOLDI, RoPS, USC, and SI descriptors. The RP curves obtained on the three datasets are shown below. Figure 3 As shown.

[0072] from Figure 3 As can be seen, on the Space Time dataset, the descriptor obtained by the method of this invention achieves a matching effect comparable to TOLDI, which is better than other descriptors. On the UWA3M and UWAOR datasets, the descriptor obtained by the method of this invention is significantly better than other descriptors.

[0073] To further understand the implementation process of the descriptors obtained by the method of this invention, a pair of point clouds is taken, and point cloud registration is performed using the descriptors obtained by the method of this invention. First, key points are extracted from the two point clouds using consistent sampling. Then, descriptors are calculated on these key points according to the method of this invention. The similarity of the descriptors is compared using the NNSR method (with a threshold set to 0.9) to establish a pair between the two point clouds. Then, a pairing clustering algorithm is used to find the correct pairing, and the transformation parameters are calculated to transform the original point cloud onto the target point cloud. The point cloud registration process is as follows: Figure 4 As shown, Figure 4 In (a), the pairings are established using the descriptors obtained by the method of this invention. Figure 4 In (b), the correct pairings were found using a pairwise clustering algorithm. Figure 4 In the diagram (c), the point cloud is after registration; the black lines in the diagram represent incorrect pairings, and the green lines represent correct pairings.

[0074] Because the descriptor of this invention has a good matching effect, it can establish more correct pairs, from Figure 4 As can be seen, the pairing clustering algorithm found many correct pairings. This is because the descriptors obtained by the method of this invention established many correct pairings, which is very beneficial to the successful registration of point clouds.

[0075] The preferred implementation of the present invention has been described in detail above, but the present invention is not limited to the described embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for obtaining binary descriptors based on local voxel encoding, characterized in that, Includes the following steps: Step S1: Calculate the local reference frame based on the local point cloud around the key point, and convert the local point cloud into the local reference frame; Step S2: Voxelize the converted local point cloud, calculate the binary string, and then concatenate all the binary strings to obtain the binary descriptor, the specific content of which is: Step S21: Convert the transformed local point cloud into voxels and divide it into 15×15×15 voxels of the same size. Let m be the occupied voxel and the other 15×15×15-m voxels be denoted as empty voxels. Step S22: From top to bottom, for each layer after local point cloud voxelization, calculate the eigenvalues ​​corresponding to all symmetric voxels, including: For two voxels that are symmetric about the main diagonal, if they are both occupied by points or are both empty voxels, then the eigenvalue is 0. If two voxels, one occupied and the other empty, are selected, then the eigenvalue is set to 1. Step S23: Combine the eigenvalues ​​calculated for all symmetric voxels into a binary string; Step S24: Ignore the voxels on the main diagonal, concatenate the binary strings of all layers to obtain the binary descriptor.

2. The binary descriptor acquisition method based on local voxel encoding according to claim 1, characterized in that, Step S1 involves calculating a local reference frame based on the local point cloud surrounding the key points and then converting the local point cloud into the local reference frame. The specific process is as follows: Extract the local point cloud around the keypoint, with the keypoint as the center and a radius of... Extracting local point clouds from circles, Set it to 15pr, where pr represents the point cloud resolution; Firstly, the key point distance is less than... Calculate the normal vectors of the points within the matrix and construct the covariance matrix. （1）； In the above formula, This indicates a distance to a key point that is less than point, This represents the average value of these points; Performing eigenvalue decomposition on the covariance matrix yields three eigenvalues ​​and three corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue is the normal vector. ; Based on the normal vector The z-axis of the local reference frame (LRF) is determined using the following formula. （2）； In the above formula, Indicate key points, This represents a point in a local point cloud. The x-axis of the local reference frame (LRF) is calculated as follows: （3）； In the above formula, It is the projection vector. and This indicates two rights; rights To improve the robustness of LRF to partial overlap, weights Improve the repeatability of LRF in flat regions; Finally, the y-axis of the local reference frame (LRF) is calculated using the cross product. （4）； The local point cloud is converted into the local reference frame (LRF) according to formulas (2), (3), and (4).

Images