A pairing selection method based on machine learning
By calculating the compatibility score between pairs and using the support vector machine SVM classifier, the problem of mispaired pairing in point cloud registration is solved, achieving higher accuracy and recall, and improving the registration success rate.
Patent Information
- Application Number
- CN202310888488.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-19
- Publication Date
- 2025-09-05
- Estimated Expiration
- 2043-07-19
AI Technical Summary
The existing point cloud registration methods have many mispairs in the case of noise, point density changes and partial overlap, resulting in low registration accuracy and long calculation time.
By calculating the compatibility score between pairs, the feature vector is constructed and classified using the support vector machine SVM classifier to find out the correct pairing.
Improve the accuracy and recall rate of point cloud registration, reduce mispair pairing, and improve the registration success rate.
Smart Images
Figure CN116842458B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of point cloud data processing, and in particular to a pairing selection method based on machine learning. Background Art
[0002] Point cloud registration is the first step in using point cloud data and plays a crucial role in many practical applications, such as scene reconstruction, 3D object recognition, and location identification. LSD-based point cloud registration has garnered widespread attention and has become a mainstream registration method. However, due to noise, point density variations, and partial overlap, LSD-based pairings often contain incorrect pairs. Therefore, finding the correct pair from the initial set of pairs is essential, making pair selection a crucial process.
[0003] Existing pairing selection methods can be roughly divided into two categories: maximum consistency (MC)-based methods and confidence verification (CV)-based methods. MC-based methods, such as geometric constraint clustering (GCC), game theoretic matching (GTM), and local and global voting (LGV), strive to identify correct pairings from an initial set of pairings. GCC utilizes geometric distance constraints to identify pairs that mutually satisfy the constraints as correct pairs. GTM also utilizes geometric distance constraints to identify compatible pairs and exclude incompatible ones. LGV uses a two-stage approach to identify correct pairs. In the local voting stage, a compatibility score is calculated for each pairing. Pairs with high compatibility scores are extracted and input into the global voting stage. In the global voting stage, a transformation parameter is calculated for each extracted pairing, and then a global vote is calculated. However, these methods still suffer from poor accuracy, with many incorrect pairings being selected and retained. CV-based methods, on the other hand, randomly select several pairs in each iteration, calculate the transformation parameters, and then verify the confidence score of the transformation parameters. The transformation with the highest confidence score is considered the optimal one. For example, the random sample consensus (RANSAC) algorithm randomly selects three pairs each time to calculate the transformation parameters. Then, based on the transformation parameters, it transforms the source point cloud to the target point cloud and calculates the number of points in the overlapping area of the two point clouds as a confidence score. The 1-point RANSAC algorithm randomly selects one pair to calculate the transformation parameters. Because it uses the local reference frame (LRF) of the key points to calculate the rotation matrix, only one pair is sufficient to calculate the transformation parameters. The 2SAC-GC (2-point based sample consensus with global constraint) algorithm randomly selects two pairs to calculate the transformation parameters and uses a geometric distance constraint and an angle constraint to reject incorrect pairs and speed up the calculation. This algorithm uses the normal vectors of the key points to calculate the rotation matrix, so two pairs are sufficient to calculate the transformation parameters. This type of algorithm cannot determine the number of iterations required and is computationally time-consuming. In addition, when there are many incorrect pairs, this type of method is prone to failure. Summary of the Invention
[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a pairing selection method based on machine learning. The compatibility score is calculated using two geometric distance constraints, and a feature vector is formed by the compatibility score. The feature vector is then input into a machine learning classifier to classify the pairings and thus find the correct pairing.
[0005] To achieve the above objectives, the present invention provides the following technical solutions.
[0006] A pair selection method based on machine learning calculates a compatibility score based on two geometric distance constraints, then constructs a feature vector based on the compatibility score. A support vector machine (SVM) classifier is then trained based on the feature vector. Finally, the trained SVM is used to predict the labels of the test data. The method includes the following steps:
[0007] Step S1, calculating the compatibility score between the pairs according to the two geometric distance constraints;
[0008] Step S2: For any pair, calculate its compatibility score with other pairs, and then construct two feature vectors. Then, use minimum pooling to combine the two feature vectors into one. Sort the elements in the feature vector from large to small, and take the first k elements as the final feature vector.
[0009] Step S3: Train a support vector machine (SVM) classifier based on the feature vectors and labels of the training data, input the feature vectors of the test data into the trained support vector machine (SVM) classifier, predict the paired labels in the test data, and find the correct pairing.
[0010] In step 1, the compatibility score between the pairs is calculated based on the two geometric distance constraints. The specific calculation process is as follows:
[0011] Assume that the initial pairing set is C = {c1 c2…c n}, for any two pairs and and are the two key points of the origin cloud, and are two key points of the target point cloud, then the first geometric distance constraint is defined as
[0012]
[0013] The second geometric distance constraint is defined as
[0014]
[0015] In the above formula, R i =V i t(V i s ) T , V i s , V i t , and The key points are and The local reference frame LRF;
[0016] Calculate two compatibility scores based on two geometric distance constraints
[0017]
[0018] In the above formula, σ1 and σ2 are two Gaussian parameters. σ1 is generally set to 1pr, and σ2 is generally set to 40pr. pr represents the point cloud resolution, that is, the average distance between all points in the point cloud and their nearest neighbor points; s1(c i ,c j ) and s2(c i ,c j ) takes a value between 0 and 1.
[0019] In step S2, for any pair, the compatibility score between it and other pairs is calculated, and then two feature vectors are constructed. The specific process is as follows:
[0020] For any pair c, calculate the compatibility score between it and all other pairs in the initial pair set, and construct two feature vectors
[0021] F1=[s1(c,c1) s1(c,c2) … s1(c,c n-1 )] (5)
[0022] F2=[s2(c,c1)s2(c,c2)…s2(c,c n-1 )] (6)
[0023] Using minimum pooling, the two feature vectors are combined to obtain
[0024] F = minpool(F1, F2) (7)
[0025] Sort the elements in F from large to small, take the first k largest elements, and get the final eigenvector
[0026]
[0027] In the above formula, the k value is generally set to 40.
[0028] In step S3, the feature vector of the test data is input into the trained support vector machine (SVM) classifier to predict the paired labels in the test data and find the correct pairing. The specific process is as follows:
[0029] First, descriptors are used to establish pairs for each pair of point clouds. Some pairs are used as training data. Based on the actual transformation relationship between the point clouds, correct pairs are determined and labeled as 1, while incorrect pairs are labeled as 0. The feature vectors of these pairs are calculated, and a support vector machine (SVM) classifier is trained based on the feature vectors and labels. The trained support vector machine (SVM) classifier is then used with the feature vectors of the test data pairs to predict the labels of the test data pairs. Pairs labeled as 1 are correct pairs. Compared with the prior art, the present invention has the following advantages:
[0030] The present invention calculates compatibility scores between pairs based on two geometric distance constraints. For any pair, its compatibility score with other pairs is calculated to form two feature vectors. These two feature vectors are then combined into one using minimum pooling. The feature vectors are sorted from largest to smallest, and the first k elements are taken as the final feature vector. An SVM classifier is then trained based on the feature vectors and labels of the training data (e.g., a correct pairing is labeled 1, an incorrect pairing is labeled 0). The feature vectors of the test data are then input into the trained SVM classifier to predict the labels of the pairs in the test data, thereby identifying the correct pairing. Compared to existing GC, SI, and CV methods, the present method can achieve higher precision and recall, identifying more correct pairs while minimizing incorrect pairs, thus improving the registration success rate. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 This is a flow chart of a pairing selection method based on machine learning of the present invention;
[0032] Figure 2 Point cloud registration result. DETAILED DESCRIPTION
[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.
[0034] Example 1
[0035] like Figure 1As shown, this embodiment provides a pair selection method based on machine learning, which calculates a compatibility score based on two geometric distance constraints, then constructs a feature vector based on the compatibility score, then trains a support vector machine (SVM) classifier based on the feature vector, and finally predicts the label of the test data based on the trained support vector machine (SVM), including the following steps:
[0036] Step S1, calculating the compatibility score between the pairs according to the two geometric distance constraints;
[0037] Step S2: For any pair, calculate its compatibility score with other pairs, and then construct two feature vectors. Then, use minimum pooling to combine the two feature vectors into one. Sort the elements in the feature vector from large to small, and take the first k elements as the final feature vector.
[0038] Step S3: Train a support vector machine (SVM) classifier based on the feature vectors and labels of the training data, input the feature vectors of the test data into the trained support vector machine (SVM) classifier, predict the paired labels in the test data, and find the correct pairing.
[0039] In step 1, the compatibility score between the pairs is calculated based on the two geometric distance constraints. The specific calculation process is as follows:
[0040] Assume that the initial pairing set is C = {c1 c2…c n}, for any two pairs and and are the two key points of the origin cloud, and are two key points of the target point cloud, then the first geometric distance constraint is defined as
[0041]
[0042] The second geometric distance constraint is defined as
[0043]
[0044] In the above formula, R i =V i t (V i s ) T , V i s , V i t , and The key points are and The local reference frame LRF;
[0045] Calculate two compatibility scores based on two geometric distance constraints
[0046]
[0047] In the above formula, σ1 and σ2 are two Gaussian parameters. σ1 is generally set to 1pr, and σ2 is generally set to 40pr. pr represents the point cloud resolution, that is, the average distance between all points in the point cloud and their nearest neighbor points; s1(c i ,c j ) and s2(c i ,c j ) takes a value between 0 and 1.
[0048] In step S2, for any pair, the compatibility score between it and other pairs is calculated, and then two feature vectors are constructed. The specific process is as follows:
[0049] For any pair c, calculate the compatibility score between it and all other pairs in the initial pair set, and construct two feature vectors
[0050] F1=[s1(c,c1) s1(c,c2) … s1(c,c n-1 )] (5)
[0051] F2=[s2(c,c1)s2(c,c2)…s2(c,c n-1 )] (6)
[0052] Using minimum pooling, the two feature vectors are combined to obtain
[0053] F = minpool(F1, F2) (7)
[0054] Sort the elements in F from large to small, take the first k largest elements, and get the final eigenvector
[0055]
[0056] In the above formula, the k value is generally set to 40.
[0057] In step S3, the feature vector of the test data is input into the trained support vector machine (SVM) classifier to predict the paired labels in the test data and find the correct pairing. The specific process is as follows:
[0058] First, descriptors are used to establish pairings for each pair of point clouds. Some of the pairings are used as training data. Based on the actual transformation relationship between the point clouds, the correct pairings are determined and labeled as 1, and the incorrect pairings are labeled as 0. The feature vectors of these pairings are calculated, and a support vector machine (SVM) classifier is trained based on the feature vectors and labels. Then, the trained support vector machine (SVM) classifier is used to predict the labels of the test data pairs using the feature vectors of the test data pairs. The pairings labeled as 1 are the correct pairings.
[0059] Example 2
[0060] The following is a specific experiment to compare the method of the present invention with traditional GC, SI and CV methods to verify the scientific nature of the algorithm of the present invention.
[0061] The experimental process of this embodiment is as follows. The experiments are carried out using the BR, UWA3M and augmented ICL-NUIM datasets. The detailed information of the three datasets is shown in Table 1 below.
[0062] Table 1 Detailed information of the three datasets
[0063]
[0064] For each pair of point clouds, keypoints are extracted using a continuous sampling algorithm with a sampling interval of 7pr (pr represents the point cloud resolution). TriLCI descriptors are then calculated and the similarity of the descriptors is compared using the NNSR algorithm to establish an initial set of pairings. A quarter of the initial set of point cloud pairs is used as training data to train a SVM classifier. The remaining pairs are used as test data, and the trained SVM classifier is used to find the correct pair.
[0065] The proposed method is compared with the GC, SI and CV methods, and the effects of each method are evaluated from the perspective of precision, recall and F score. The three indicators are defined as
[0066]
[0067] In the above formula, m1 is the number of correct pairs found, m2 is the number of pairs found, and m3 is the number of correct pairs in the initial pairing set. The calculation results of the four methods are shown in Table 2 below.
[0068] Table 2 Precision, recall and F-score obtained by four pair selection methods
[0069]
[0070]
[0071] The results in Table 2 show that on the BR dataset, our method achieved the best recall and F-score, and the second-best precision. On the UWA3M and Augmented ICL-NUIM datasets, our method achieved the best precision and F-score. For point cloud registration, registration accuracy is even more important. Therefore, overall, our method outperforms the other three methods.
[0072] Next, the method of the present invention is used in point cloud registration, and the registration is performed on three pairs of point clouds. The results are as follows: Figure 2 shown.
[0073] exist Figure 2 In the figure, each sub-figure shows the registration process of a pair of point clouds. The initial point cloud position and the initial pairing set are shown on the left, the pairing found by the method of the present invention is shown in the middle figure, and the registered point cloud is shown on the right. The black line in the figure indicates an incorrect pairing, and the green line indicates a correct pairing. Figure 2 It can be seen that the method of the present invention can find the correct pairing very well.
[0074] The above specifically describes the preferred implementation method of the present invention, but the invention is not limited to the described embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.
Claims
1. A pair selection method based on machine learning, which calculates a compatibility score based on two geometric distance constraints, then constructs a feature vector based on the compatibility score, then trains a support vector machine (SVM) classifier based on the feature vector, and finally predicts the label of the test data based on the trained support vector machine (SVM), characterized by: The following steps are involved: Step S1, calculating the compatibility score between the pairs according to the two geometric distance constraints; Assume that the initial pairing set is C = {c1 c2…c n }, for any two pairs and and are the two key points of the origin cloud, and are two key points of the target point cloud, then the first geometric distance constraint is defined as The second geometric distance constraint is defined as In the above formula, and The key points are and The local reference frame LRF; Calculate two compatibility scores based on two geometric distance constraints In the above formula, σ1 and σ2 are two Gaussian parameters. σ1 is generally set to 1pr, and σ2 is generally set to 40pr. pr represents the point cloud resolution, that is, the average distance between all points in the point cloud and their nearest neighbor points; s1(c i ,c j ) and s2(c i ,c j ) takes a value between 0 and 1; Step S2: For any pair, calculate its compatibility score with other pairs, and then construct two feature vectors. Then, use minimum pooling to combine the two feature vectors into one. Sort the elements in the feature vector from large to small, and take the first k elements as the final feature vector. Step S3: Train a support vector machine (SVM) classifier based on the feature vectors and labels of the training data, input the feature vectors of the test data into the trained support vector machine (SVM) classifier, predict the paired labels in the test data, and find the correct pairing.
2. A pairing selection method based on machine learning according to claim 1, characterized in that: In step S2, for any pair, the compatibility score between it and other pairs is calculated, and then two feature vectors are constructed. The specific process is as follows: For any pair c, calculate the compatibility score between it and all other pairs in the initial pair set, and construct two feature vectors F1=[s1(c,c1) s1(c,c2) … s1(c,c n-1 )] (5) F2=[s2(c,c1)s2(c,c2)…s2(c,c n-1 )] (6) Using minimum pooling, the two feature vectors are combined to obtain F = minpool(F1, F2) (7) Sort the elements in F from large to small, take the first k largest elements, and get the final eigenvector In the above formula, the k value is generally set to 40.
3. The pairing selection method based on machine learning according to claim 1, characterized in that: In step S3, the feature vector of the test data is input into the trained support vector machine (SVM) classifier to predict the paired labels in the test data and find the correct pairing. The specific process is as follows: First, descriptors are used to establish pairings for each pair of point clouds. Some of the pairings are used as training data. Based on the actual transformation relationship between the point clouds, the correct pairings are determined and labeled as 1, and the incorrect pairings are labeled as 0. The feature vectors of these pairings are calculated, and a support vector machine (SVM) classifier is trained based on the feature vectors and labels. Then, the trained support vector machine (SVM) classifier is used to predict the labels of the test data pairs using the feature vectors of the test data pairs. The pairings labeled as 1 are the correct pairings.