A low-overlap-rate point cloud registration method for autonomous navigation of a drone

Abstract

The present application relates to a kind of low overlap rate point cloud registration methods for unmanned aerial vehicle autonomous navigation, comprising: point cloud feature score matrix is as cost matrix, score matrix is optimized in optimization iteration in transmission process, so as to reduce the redundant interaction between features, guarantee the distinguishability of feature;By overlapping degree score and significant score in priori prediction to each point in point cloud pair, the score of point is set as the target cost of optimization transmission, so as to realize controllable attention interaction under the guidance of priori score;By the initial matching relationship in cost matrix, geometric consistency screening is carried out, and high score error matching pair in cost matrix is effectively corrected.Finally, based on the proposed sparse feature interaction, priori score guide, space consistency regularization, a point cloud registration network from coarse matching to fine matching is designed, for establishing matching point pair and estimating point cloud relative transformation, can be used in unmanned aerial vehicle navigation, virtual reality, three-dimensional reconstruction and multiple related fields.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and particularly relates to a low overlap ratio point cloud registration method for autonomous navigation of a UAV. BACKGROUND

[0002] The goal of point cloud registration is to compute the pose transformation between two partially overlapping point clouds in the same scene to align the point clouds. Spatial position perception is a core part of the autonomous navigation process of a UAV, and is the basis for the UAV to plan a correct navigation path according to the surrounding environment and its own position. Point cloud registration can provide real-time pose for the UAV according to scanned point cloud data. With the development and application of deep learning technology, more and more works introduce deep learning methods into point cloud feature coding, matching point establishment and pose solving. Such methods use the coding ability of deep learning to encode geometric structure information into the features of each point or use the fitting ability of the network to achieve fast pose calculation. Using attention mechanisms to fuse point cloud features has also become a basic paradigm for constructing point cloud features today. However, these methods still have shortcomings in feature construction: 1) the existing attention mechanism realizes the dense connection of all features of the source point cloud and the target point cloud, which will cause a large number of redundant feature interaction connections between the point clouds, which will cause the features to be matched to gradually blur in the interaction process. Excessively smoothed features will produce a large number of false matches in the matching stage, and thus bring about false pose estimation. 2) The existing methods all use point cloud features for direct enhancement processing, and the features of the outliers will interfere with the features of the inliers and affect the distribution of the final features, so that there is often a situation of patch false match aggregation. 3) The existing methods often cannot normally register the low overlap ratio point clouds scanned by the UAV when flying at high speed. SUMMARY

[0003] The present application provides a low overlap ratio point cloud registration method for autonomous navigation of a UAV, which solves the technical problems that the existing feature construction methods are prone to feature blurring and low overlap ratio point clouds are difficult to register.

[0004] According to a first aspect of the present application, a low overlap ratio point cloud registration method for autonomous navigation of a UAV is provided, comprising:

[0005] Step 1, using a multi-layer down-sampling method based on kernel point convolution to down-sample the input partially overlapping source point cloud X and target point cloud Y, performing structured feature extraction on point clouds of different scales in each layer through a three-dimensional convolution operation to obtain the source point cloud node set XC and the target point cloud node set YC at the coarsest scale and the corresponding node features at the coarsest scale and

[0006] Step 2, the node features are fused by using a feature attention fusion and overlap score prediction method And The fused node features are obtained by performing feature fusion And the overlap scores of each node are regressed And The overlap scores And are probabilities of each node being located in an overlapping area of the point cloud;

[0007] Step 3, according to the distinguishability of the node features , the similarity of each node feature to all node features in a neighborhood range is calculated, and the saliency scores of each node are calculated by proportion And The saliency scores And are the uniqueness degrees of each node feature in the neighborhood range;

[0008] Step 4, global feature enhancement and cross-point cloud feature fusion are performed on the node features based on a self-attention-cross-attention mechanism;

[0009] Step 5, the scores of the attention matrix in step 4 are taken as matching scores, a point cloud matching pair is constructed, the matching scores of the matching pair are dynamically adjusted by using the spatial consistency of the matching pair, and a feature interaction relationship conforming to the spatial consistency is formed;

[0010] Step 6, the matching relationship is constructed based on the fused features, the point cloud block features are matched, and a node pair represented by a high-score is selected, and a matching relationship at a coarse scale is formed;

[0011] Step 7, a multi-layer upsampling operation is performed on the original node features, the down-sampled point cloud is restored to the original scale, and fine features of each point are obtained, fine point features are matched in each block matching represented by the coarse matching relationship formed in step 6, a confidence score of each fine point pair is saved, all fine matching relationships and confidence scores are ranked, and a fine matching with a high ranking is selected as a final matching point pair;

[0012] Step 8, based on the matching point pair obtained in step 7, a relative pose transformation between the two point clouds is estimated by using a RANSAC method.

[0013] On the basis of the above technical solutions, the application can also be improved as follows.

[0014] Optionally, in step 1, the input source point cloud The target point cloud

[0015] with are the source and target point cloud node sets at the coarsest scale after down-sampling process, respectively, N C and M C represent the number of points in the source and target point cloud at this scale, respectively, D C represents the dimension of the point cloud feature at this scale.

[0016] The point cloud points at fine scales are constructed using the neighbor search to construct the neighbor relationship between the point cloud points at fine scales and the final down-sampled points, and a local point cloud block is constructed for each down-sampled point using the coarsest scale feature characterizes the point cloud block features in the source and target point clouds.

[0017] Optionally, the step 2 comprises:

[0018] Step 201, using the original feature attention module to perform self-attention and cross-attention fusion on the node features ;

[0019] Step 202, the adopted attention layer stacking mode is a 5-layer cross structure of self-cross-self-cross-self, which ensures that the features used for prior prediction have a receptive field in the global range, and the obtained features are subjected to overlap score regression using a fully connected network;

[0020] Step 203, the neighbor relationship is calculated for the point cloud before sampling and the point cloud after sampling, the neighborhood original points corresponding to each node are found, the spatial position of each node under the real matching is calculated using the real pose transformation matrix, the point cloud overlap region is calculated and the ratio of the original point in the overlap region where each node is located is judged, the overlap degree of each node is determined by the percentage of the overlap region part in the node neighborhood area and is taken as the final real overlap score

[0021] Step 204, the overlap score obtained by regression is normalized and adjusted, through truncation operation, the nodes with predicted overlap score lower than the set truncation threshold are regarded as non-overlapping points, and the overlap score is corrected to 0; the nodes with predicted overlap score higher than the set truncation threshold are regarded as overlapping points, and the overlap score is corrected to 1.

[0022] Optionally, the process of cross-attention fusion in the step 201 comprises:

[0023]

[0024] wherein Linear is a linear mapping, representing the mapping of the node features Mapping features to be used for attention mechanism operation

[0025] Using softmax on similarity matrix To get cost matrix w;

[0026] Cost matrix w and features Weighted reconstruction and original features Superimposed, the superimposed features are mapped using a fully connected network MLP, and the original node features To construct the final fusion features

[0027] Optionally, the step 3 includes:

[0028] Comparing node features and neighborhood feature similarity and statistics to obtain the significance score of each node formula:

[0029]

[0030] Where, norm is the feature normalization operation; The neighborhood points of the node Each node's neighborhood points include the node itself; s thre The set of significance threshold; i and j represent the serial number of the node.

[0031] Optionally, the step 4 includes:

[0032] Step 401, the feature similarity score matrix is used as the cost matrix C of optimization transmission, a row of garbage bin constraint row and a column of garbage bin constraint column are added to the cost matrix C, and the cost matrix Augmented to

[0033]

[0034] The augmented cost matrix Replace the original cost matrix C, and obtain a new result matrix

[0035] Step 402, the overlap score and the significance score are combined as the row constraint μ r And column constraint μ c Of the cost matrix.

[0036] Step 403, using the optimization transmission iteration equation to update the cost matrix iteratively:

[0037]

[0038] where the final optimized target matrix w OT is the resulting matrix is a sub-matrix of

[0039] The iteration aims to maximize the cost equation

[0040] Optionally, the step 5 comprises:

[0041] Step 501, for the sparse attention matrix of step 4, select representative high-score points as the hypothesis matching relationship, and find the corresponding points in the three-dimensional space according to the index of the points; perform spatial consistency score calculation on the hypothesis matching relationship:

[0042]

[0043] where [·] + is the max(·,0) operation; σ d represents the distance threshold; ||·|| is the L2-Norm, and |·| is the absolute value function;

[0044] Step 502, update the cost score on the cost matrix using the spatial consistency scores SC ij of all hypothesis point pairs:

[0045]

[0046] where λ represents the smoothing weight, and T represents the set of assumed matches.

[0047] Optionally, the step 6 comprises:

[0048] Step 601, use the optimized node features obtained in step 5 to perform feature matching, each node of the source point cloud is compared with the features of each node of the target point cloud for similarity, and the original Sinkhorn method is used to iteratively optimize the score matrix to obtain a final coarse matching score matrix;

[0049] Step 602, the coarse matching score matrix reflects N c ×M c matching relationships, and 256 items with higher scores are selected from the matching relationships as matching relationships at a coarse scale.

[0050] Optionally, in step 7, a local fine matching set is constructed for each coarse match, and the confidence score corresponding to each fine match is saved. After integrating all fine matches corresponding to the coarse matches, the fine match pairs with higher scores are selected as the final matching point pairs according to the score sorting.

[0051] Optionally, the step 8 comprises:

[0052] Step 801, randomly select 3 matching pairs from the final matching pairs obtained in the step 7 as RANSAC sampling matching, and solve the rigid transformation matrix according to the selected three matching pairs;

[0053] Step 802, calculate the distance error of the remaining point pairs in the final matching pairs under the rigid transformation matrix; if the distance error of a matching pair is less than a set error threshold, the point is an inlier, otherwise it is an outlier; count the proportion of inliers in all points;

[0054] Step 803, repeat steps 801 and 802 until the upper limit of the iteration number, and select the rigid transformation corresponding to the highest inlier proportion in the iteration process as the final estimated point cloud relative pose.

[0055] The application provides a low-overlap point cloud registration method for autonomous navigation of a UAV, proposes a general sparse attention construction method, iteratively optimizes the cost matrix by using optimized transmission, can effectively reduce the redundant connection of point cloud features when interacting across point clouds, and effectively avoids the excessive smoothing and blurring of point cloud features; an attention interaction mode guided by prior information is proposed, the existing prior information can be used to explicitly constrain the optimized transmission iteration process, so that only features in the overlapping part of the point cloud and the area with distinguishability interact, the imbalance of feature distribution in the case of low overlap rate point cloud is improved, and the positioning accuracy of the UAV in high-speed flight is improved. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 A flowchart of a low-overlap point cloud registration method for autonomous navigation of a UAV is provided for the embodiments of the application. DETAILED DESCRIPTION

[0057] The principles and characteristics of the application are described below in combination with the drawings, and the examples are only used to explain the application and not to limit the scope of the application.

[0058] Figure 1 A flowchart of a low-overlap point cloud registration method for autonomous navigation of a UAV is provided for the embodiments of the application. Figure 1 As shown in the figure, the registration method comprises:

[0059] Step 1, use a multi-layer down-sampling method based on kernel point convolution to down-sample the input partially overlapping source point cloud X and target point cloud Y, perform structured feature extraction on point clouds of different scales in each layer through three-dimensional convolution operation, and obtain source point cloud node set X C and target point cloud node set YC and its corresponding node feature of the coarsest scale and

[0060] Step 2, using the feature attention fusion and overlap score prediction method to fuse the node features and to obtain the fused node features and regress the overlap score of each node and overlap score and the probability of each node being in the overlapping area of the point cloud.

[0061] Step 3, according to the distinguishability of the node features , calculate the similarity of each node feature with all node features in the neighborhood range, and obtain the saliency score of each node by proportion calculation and saliency score and the uniqueness of each node feature in the neighborhood range.

[0062] Step 4, based on the self-attention-cross-attention mechanism, perform global feature enhancement and cross-point cloud feature fusion on the node features.

[0063] Step 5, take the score of the attention matrix in step 4 as the matching score, construct the point cloud matching pair, and use the spatial consistency of the matching pair to dynamically adjust the score of the matching pair, so that the score in the cost matrix tends to the correct distribution, thereby forming a feature interaction relationship that conforms to the spatial consistency.

[0064] Step 6, based on the fused features, construct the matching relationship, perform feature matching on the point cloud block features, and select the node pairs represented by the high-score scores to form the matching relationship at the coarse scale.

[0065] Step 7, use the original node features to perform multi-layer upsampling operation, restore the down-sampled point cloud to the original scale, and obtain the fine features of each point. In the block matching represented by each coarse matching of the matching relationship at the coarse scale formed in step 6, perform fine point feature matching, save the confidence score of each fine point pair, integrate all fine matching relationships and confidence scores for ranking, and select the top-ranking fine matching as the final matching point pair.

[0066] Step 8, based on the matching point pair obtained in step 7, estimate the relative pose transformation between the two point clouds using the RANSAC method.

[0067] The application provides a low-overlap point cloud registration method for autonomous navigation of a UAV.

[0068] Embodiment 1

[0069] The embodiment 1 provided by the application is an embodiment of the low-overlap point cloud registration method for autonomous navigation of a UAV provided by the application, which combines Figure 1 It can be seen that the embodiment of the registration method comprises:

[0070] Step 1, using a multi-layer down-sampling method based on kernel point convolution to down-sample the input partially-overlapping source point cloud X and target point cloud Y, performing structured feature extraction on the point cloud of each layer with different scales through a three-dimensional convolution operation to obtain the source point cloud node set X C and the target point cloud node set Y C and their corresponding node features of the coarsest scale and

[0071] In a possible embodiment, the input source point cloud target point cloud

[0072] and are respectively the source point cloud node set and the target point cloud node set of the coarsest scale obtained through the down-sampling process, N C and M C respectively represent the number of points in the source point cloud and the target point cloud at this scale, and D C represents the dimension of the point cloud feature at this scale.

[0073] The point cloud points in the fine scale are used to construct the neighbor relationship of the final down-sampled points by using the neighbor search, and a local point cloud block is constructed for each down-sampled point; the coarsest scale feature is used to represent the point cloud block feature in the source point cloud and the target point cloud.

[0074] Step 2, using a feature attention fusion and overlap score prediction method to perform feature fusion on the node features and to obtain the fused node feature and regress the overlap score of each node and The overlap score and is the probability of each node being located in the overlapping area of the point cloud.

[0075] In a possible implementation mode, step 2 comprises:

[0076] Step 201, using the original feature attention module to perform self-attention and cross-attention fusion on the node features .

[0077] In a possible implementation mode, the process of performing cross-attention fusion in step 201 comprises:

[0078]

[0079] wherein Linear is a linear mapping, representing mapping of the node features to features used for attention mechanism operation. In a specific implementation, for the self-attention mechanism, the three input features therein are replaced by or The remaining parts are the same structure as the cross-attention mechanism.

[0080] The similarity matrix is normalized using softmax to obtain the cost matrix w.

[0081] The cost matrix w is weighted and reconstructed with the features and is superimposed with the original features , the superimposed features are mapped using a fully connected network MLP, and are used together with the original node features to construct the final fusion features

[0082] Step 202, the attention layer adopted is a 5-layer cross structure of self-cross-self-cross-self, which ensures that the features used for prior prediction have a global receptive field, and the obtained features are used for overlap score regression using a fully connected network.

[0083] Step 203, the neighbor relationship is calculated for the point cloud before sampling and the point cloud after sampling, the original points corresponding to each node are found, and the real pose transformation matrix T gt is used to calculate the spatial position of each node under real matching, the point cloud overlap region is calculated and the ratio of the original point of each node located in the overlap region is judged, the overlap degree of each node is determined by the percentage of the overlap region part in the node neighborhood area, and the real overlap score is taken as the final real overlap score

[0084] Step 204, normalize the regression obtained overlap score, through the truncation operation, the nodes with predicted overlap score lower than the set truncation threshold are regarded as non-overlapping points, and the overlap score is corrected to 0; the nodes with predicted overlap score higher than the set truncation threshold are regarded as overlapping points, and the overlap score is corrected to 1. In a specific implementation, the truncation threshold can be set to 0.02.

[0085] Step 3, according to the distinguishability of the node features , the similarity of each node to all node features within the neighborhood range is calculated, and the significance score of each node is calculated by the proportion and the significance score and is the uniqueness of each node feature within the neighborhood range.

[0086] In one possible embodiment, step 3 includes:

[0087] The significance score calculation process of each node is based on feature similarity comparison, and the formula for calculating the significance score of each node by comparing the similarity of node features and neighborhood features is:

[0088]

[0089] where norm is the feature normalization operation; is the neighborhood point of the node , and the neighborhood point of each node includes the node itself; s thre is the set significance threshold; i and j represent the serial number of the nodes respectively.

[0090] In a specific implementation, the similarity of each node feature and each neighbor feature is calculated, and the total number of feature pairs with similarity greater than the significance threshold is counted, and the significance threshold can be set to 0.6. The more the number is, the more node features similar to the node feature within a certain range, that is, the worse the distinguishability of the node feature is. Finally, the significance score of each node is calculated using the reciprocal.

[0091] Considering that each node and its neighborhood point includes the node itself, the number of nodes with similarity greater than the similarity threshold in node similarity calculation is at least 1, so the reciprocal statistics of the significance score will not cause division by zero problem.

[0092] Step 4, based on the self-attention-cross-attention mechanism, the node features are globally enhanced and cross-point cloud features are fused.

[0093] In the implementation, step 4 is in the feature cross fusion stage, the feature similarity score matrix is taken as the cost matrix of the optimization transmission, the overlap score and the saliency score obtained in steps 2 and 3 are taken as the constraint term of the optimization transmission, the out-point score in the non-overlapping area is controlled by using the cost score garbage can mechanism, and the attention matrix in the cross attention mechanism is iteratively optimized based on the optimization transmission method.

[0094] In one possible implementation manner, step 4 includes:

[0095] In step 401, the feature similarity score matrix is taken as the cost matrix C of the optimization transmission, a row of garbage can constraint rows and a column of garbage can constraint columns are added to the cost matrix C, and the cost matrix C is augmented to

[0096]

[0097] The augmented cost matrix C is replaced with the original cost matrix C, and a new result matrix w is obtained after iteration by using the Sinkhorn algorithm.

[0098] Since the row sum and the column sum are constrained by μ r and μ c , the larger the garbage can value of a certain row or column is, the smaller the row sum or column sum of w OT will be, which means that the algorithm can implicitly control the interaction of the point corresponding to the row or column through the garbage can, and the algorithm allows the network to learn the ability to interact more with more important points.

[0099] In step 402, the overlap score and the saliency score are jointly taken as the row constraint μ r and the column constraint μ c of the cost matrix, so the distribution of the cost score in the Sinkhorn iteration process can be controlled in advance, the cost score of the area with higher overlap and the area with stronger saliency is higher, and therefore the center of interaction is more inclined to the overlapping area and the area with large distinguishability in the feature interaction process.

[0100] In step 403, the sparse attention method guided by the optimization transmission uses the optimization transmission iteration equation to iteratively update the cost matrix:

[0101]

[0102] Wherein, the final optimization target matrix w OT is a sub-matrix of the result matrix w . ​​

[0103] The iteration target is to maximize the cost equation under the constraint of the constraint term on the row and column of the cost matrix The sparse cost matrix can be finally obtained by Sinkhorn algorithm iteration for row cost and column cost. In a specific implementation, the number of iterations of the used iteration optimization is 30.

[0104] Step 5, taking the score of the attention matrix in step 4 as a matching score, constructing a point cloud matching pair, and dynamically adjusting the score of the matching pair by using the spatial consistency of the matching pair, so that the score in the cost matrix tends to be more correct distribution, thereby forming a feature interaction relationship consistent with spatial consistency.

[0105] In a possible implementation manner, step 5 comprises:

[0106] Step 501, for the sparse attention matrix in step 4, selecting a representative high-score point as a hypothetical matching relationship, and finding its corresponding point in the three-dimensional space according to the index of the point; the three-dimensional rigid body matching point satisfies the distance consistency relationship, that is, for two pairs of fixed matching pairs, the distance between the end points on the same point cloud is consistent, and this property will be met when the point cloud is rotated or translated; therefore, the spatial consistency score of the hypothetical matching relationship is calculated:

[0107]

[0108] Wherein, [·] + max(·,0) operation is used to ensure that the consistency score is non-negative; σ d represents the distance threshold; ||·|| is L2-Norm, and |·| is the absolute value function.

[0109] Therefore, the spatial consistency score SC ij of the hypothetical point pair satisfying the consistency condition is higher, and the spatial consistency score SC ij of the hypothetical point pair not satisfying the condition is close to 0.

[0110] Step 502, using the spatial consistency score SC ij of all hypothetical point pairs to update the cost score on the cost matrix:

[0111]

[0112] Wherein, λ represents a smoothing weight, and in a specific implementation, the smoothing weight can be set to 0.6, and T represents a set of assumed matching.

[0113] According to the above formula, the cost score is updated, and if two hypothetical matching pairs have a higher consistency score SC ijThen their cost smooth weight will be higher, so as to further optimize the distribution of the cost matrix.

[0114] Step 6, based on the fused features, the matching relationship is constructed. The point cloud block features are matched, and the node pairs with high scores are selected to form the matching relationship at a coarse scale.

[0115] In one possible implementation manner, step 6 includes:

[0116] Step 601, the optimized node features obtained in step 5 are used for feature matching. Each node of the source point cloud is compared with the features of each node of the target point cloud in terms of similarity, and the original Sinkhorn method is used to iteratively optimize the score matrix to obtain a final coarse matching score matrix.

[0117] Step 602, the coarse matching score matrix reflects N c ×M c match relationships. From the matching relationships, 256 items with higher scores are selected as the final result of coarse matching, i.e., the matching relationship at a coarse scale.

[0118] Step 7, the original node features are used for multi-layer upsampling operation to restore the down-sampled point cloud to the original scale and obtain fine features of each point. Fine point features are matched in each block matching represented by the coarse matching of the matching relationship at a coarse scale formed in step 6. The confidence scores of each fine point pair are saved. All fine matching relationships and confidence scores are ranked, and the top-ranked fine matching is selected as the final matching point pair.

[0119] In one possible implementation manner, in step 7, a local fine matching set is constructed for each coarse matching, and the confidence scores corresponding to each fine matching are saved. After integrating the fine matchings corresponding to all coarse matchings, the fine matchings with higher scores are selected as the final matching point pairs according to the score ranking.

[0120] Step 8, based on the matching point pairs obtained in step 7, the RANSAC method is used to estimate the relative pose transformation between the two point clouds.

[0121] In one possible implementation manner, step 8 includes:

[0122] Step 801, three matching pairs are randomly selected from the final matching pairs obtained in step 7 as RANSAC sampling matches, and a rigid transformation matrix is solved according to the selected three matching pairs.

[0123] Step 802, calculate the distance error of the remaining point pairs in the final matching pair under the rigid transformation matrix; if the distance error of a certain matching pair is less than the set error threshold, the point is an inlier, otherwise it is an outlier; count the proportion of inliers in all points in this estimation.

[0124] Step 803, repeat steps 801 and 802 until the upper limit of the iteration number is reached, and select the rigid transformation corresponding to the iteration with the highest inlier proportion in the iteration process as the final estimated point cloud relative pose. In a specific implementation, the iteration number can be set to 50000, and the distance error threshold can be set to 0.05m.

[0125] The low-overlap point cloud registration method for autonomous navigation of the unmanned aerial vehicle provided by the embodiment of the application has significant gains in the point cloud registration result, and the gains mainly come from the following three aspects: first, an effective attention sparsity method is proposed, so that redundant feature interaction can be avoided when the point cloud features are enhanced by cross-attention, thereby effectively avoiding the problem that the features become blurred in the attention stage, and the distinguishability of the point cloud features is enhanced. Second, a priori guided attention connection method is proposed, which introduces simple prior information into the constraint term of optimization transmission, so that the cost distribution of feature matching is more inclined to the overlapping area and the feature salient area. Finally, the spatial consistency information is used to correct the connection relationship in the feature interaction process, which well avoids the error guidance of the features in the interaction process, and ensures that the unmanned aerial vehicle has good matching effect and positioning accuracy for low-overlap point clouds in flight.

[0126] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0127] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0128] The present application is described in reference to the flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0129] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0130] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0131] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those of skill in the art once they have the benefit of the present disclosure without departing from the spirit and scope of the application. Accordingly, the attached claims are intended to cover all such variations and modifications as falling within the scope of the application.

[0132] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A low overlap ratio point cloud registration method for autonomous navigation of a UAV, characterized in that, The point cloud registration method comprises: Step 1, down-sampling the input partial overlapping source point cloud X and target point cloud Y by using a multi-layer down-sampling method based on kernel point convolution, and extracting structured features of each layer of point cloud with different scales by three-dimensional convolution operation to obtain the source point cloud node set X at the coarsest scale C and the target point cloud node set Y C and their corresponding node features at the coarsest scale and Step 2, the node features are fused by using a feature attention fusion and overlap score prediction method and The fused node features are obtained by feature fusion and the overlap score of each node is regressed and The overlap score and is the probability that each node is located in the overlapping area of the point cloud Step 3, according to the distinguishability of the node features , the similarity of each node with all node features in the neighborhood range is calculated, and the prominence score of each node is calculated by the proportion and The prominence score and is the uniqueness of each node feature in the neighborhood range; Step 4, global feature enhancement and cross-point cloud feature fusion are performed on the node features based on a self-attention-cross-attention mechanism; Step 5, scores of the attention matrix of step 4 are taken as matching scores, a point cloud matching pair is constructed, a dynamic adjustment is performed on the matching pair scores by using spatial consistency of the matching pair, and a feature interaction relationship meeting the spatial consistency is formed; Step 6, a matching relationship is constructed based on the fused features, feature matching is performed on the point cloud block features, and a node pair represented by a high-score score is selected, and a matching relationship at a coarse scale is formed; Step 7, a multi-layer upsampling operation is performed on the original node features, the down-sampled point cloud is restored to the original scale, and fine features of each point are obtained, fine point features are matched in each block matching represented by the coarse matching relationship formed in step 6, a confidence score of each fine point pair is saved, all fine matching relationships and confidence scores are ranked, and a fine matching with a high rank is selected as a final matching point pair; Step 8, based on the matching point pair obtained in step 7, a relative pose transformation between the two point clouds is estimated by using a RANSAC method.

2. The method of claim 1, wherein, In step 1, the input source point cloud The target point cloud And Respectively, the source point cloud node set and the target point cloud node set at the final coarsest scale obtained by the downsampling process, N C And M C Respectively, the number of points in the source point cloud and the target point cloud at this scale, D C Indicates the dimension of the point cloud feature at this scale; Constructing local point cloud blocks for each down-sampled point using a nearest neighbor search to construct the nearest neighbor relationships of the point cloud points within the fine scale to the final down-sampled points; using the coarsest scale features Characterizing point cloud block features in the source point cloud and the target point cloud.

3. The method of claim 1, wherein, The step 2 comprises: Step 201, using the original feature attention module to the node feature Self-attention and cross-attention fusion are performed; Step 202, the adopted attention layer stacking mode is a 5-layer cross structure of self-cross-self-cross-self, the features used for prior prediction have a receptive field in a global range, and a full connection network is used to perform overlap degree score regression on the obtained features; Step 203, calculate the near-neighbor relationship of the point cloud before sampling and the point cloud after sampling, find the original point corresponding to each node in the neighborhood, calculate the spatial position of each node under the real matching by using the real pose transformation matrix, calculate the overlapping area of the point cloud and judge the ratio of the original point in the overlapping area where each node is located, determine the overlapping degree of each node through the percentage of the overlapping area part in the node neighborhood area and take it as the final real overlap score Step 204, the regression obtained overlap degree score is normalized and adjusted, through a truncation operation, nodes with a predicted overlap degree score lower than a set truncation threshold are regarded as non-overlapping points, and the overlap degree score is corrected to 0; nodes with a predicted overlap degree score higher than the set truncation threshold are regarded as overlapping points, and the overlap degree score is corrected to 1.

4. The method of claim 3, wherein, The process of cross-attention fusion in step 201 comprises: where Linear is a linear mapping representing mapping the node features into features used for attention mechanism operations using softmax on the similarity matrix w is the cost matrix obtained by normalizing Cost matrix w and features Weighted reconstruction is performed and the original features are superimposed, and the superimposed features are mapped using a fully connected network MLP, and the original node features are constructed into the final fusion features 5. The method of claim 1, wherein, The step 3 comprises: A formula for comparing node features with neighborhood feature similarity and statistically obtaining a saliency score of each node is: wherein norm is a feature normalization operation; is a neighbor point of the node , and the neighbor points of each node include the node itself; thre is a set saliency threshold; and i and j represent the serial numbers of the nodes, respectively.

6. The method of claim 1, wherein, The step 4 comprises: Step 401, taking the feature similarity score matrix as the cost matrix C of optimization transmission, adding a row of trash bin constraint row and a column of trash bin constraint column to the cost matrix C, and augmenting the cost matrix C to to The augmented cost matrix The original cost matrix C is replaced by a new result matrix obtained after iteration by Sinkhorn algorithm Step 402, combine the overlap score and the saliency score as a row constraint μ of the cost matrix r with column constraints μ c ; Step 403, an optimized transmission iteration equation is used to iteratively update the cost matrix: where the final optimized target matrix w OT is the resulting matrix sub-matrix of The iteration targets are to maximize the cost equation under the constraints of the restriction terms 7. The method of claim 1, wherein, The step 5 comprises: Step 501, for the sparse attention matrix of step 4, a representative high-score point is selected as a hypothetical matching relationship, and corresponding points of the point in a three-dimensional space are found according to an index of the point; spatial consistency score calculation is performed on the hypothetical matching relationship: in,[·] + For the max(·,0) operation; σ d This represents the distance threshold; ||·|| is the L2-Norm, and |·| is the absolute value function; Step 502, divide SC by the spatial consistency of all the pairs of hypotheses ij Update the cost scores on the cost matrix: Wherein, λ represents a smoothing weight, and T represents a hypothetical matching set.

8. The method of claim 1, wherein, The step 6 comprises: Step 601, feature matching is performed on the optimized node features obtained in step 5, each node of a source point cloud is compared with features of each node of a target point cloud, and an original Sinkhorn method is used to iteratively optimize a score matrix, and a final coarse matching score matrix is obtained; Step 602, the coarse matching score matrix reflects N c x M c matching relationships, and 256 items with the highest scores are selected from the matching relationships as the matching relationships at the coarse scale.

9. The method of claim 1, wherein, The step 7 saves the confidence score corresponding to each fine match, and after integrating all the fine matches corresponding to the coarse matches, the fine match pairs with higher scores are selected as the final match pairs according to the scores.

10. The method of claim 1, wherein, The step 8 includes: Step 801, three match pairs are randomly selected from the final match pairs obtained in the step 7 as RANSAC sampling matches, and a rigid transformation matrix is solved according to the three selected match pairs; Step 802, the distance error of the remaining point pairs in the final match pairs under the rigid transformation matrix is calculated; if the distance error of a match pair is less than a set error threshold, the point is an inlier, otherwise, the point is an outlier; the proportion of the inliers in all points is counted; Step 803, steps 801 and 802 are repeatedly executed until the upper limit of the iteration times, and the rigid transformation corresponding to the iteration process in which the inlier proportion is highest is selected as the final estimated point cloud relative pose.

Images