Research On 3D Point Cloud Registration Technology Based On Deep Learning

With the continuous progress and development of three-dimensional scanning equipment,point cloud registration has a broad application prospect.However,in actual point cloud registration,In the presence of noise,singular values,and partial point cloud loss,the conventional point cloud registration methods prove to be inadequate.The deep learning-based point cloud alignment extracts more detailed features of point clouds through convolutional networks,which makes it robust to the above situations,but the accuracy of alignment needs to be improved.Hence,the focus of this paper is on investigating the point cloud registration method utilizing deep learning,and introducing two novel approaches,namely the point cloud registration method driven by multi-feature extraction and matching matrix,and the partial point cloud registration method driven by multi-feature extraction and self-attention mechanism.The primary contributions of this paper can be summarized as follows:(1)A multi-feature extraction and match-matrix-driven point cloud registration algorithm is proposed,aiming at the fact that the timing of three-dimensional point cloud is easily affected by noise,singular value and other factors.Firstly,A module for multi-feature extraction is proposed in this study.Initially,the module utilizes the dynamic graph convolutional network(DGCNN)to extract local geometric features from the point cloud,and then calculates the significance score of the local geometric features through the multi-layer perceptron,and obtains the spatial features of the point cloud according to the significance score.Then,the proposed features are separately calculated by the matching matrix to get the respective matching matrix,and the two matching matrices are summed to get the matching point pairs,which improves the probability of correctly matching the point pairs;then the two matching matrices are aggregated and calculated by the weights to get the weights of the point pairs.Finally,the weights and matched point pairs are decomposed by weighted singular values to obtain a rigid transformation,which reduces the influence of noise and singular values and other factors on the alignment and improves the alignment accuracy.The results of ablation experiment and comparison experiment show that,compared with traditional point cloud registration methods,the proposed method has higher registration accuracy and stronger robustness in the presence of noise and singular values.(2)To address the issue of decreased registration accuracy caused by the absence of partial point clouds,we propose a novel partial point cloud registration method driven by multi-feature extraction and self-attention mechanism.Firstly,dynamic convolutional network(DGCNN)is replaced by adaptive convolution network(Adapt Conv)to extract partial point cloud features.Then,the hybrid features of the two point clouds are extracted by the cross-attention mechanism and the corresponding point pairs with robustness are inferred.Finally,the proposed method utilizes the feature correlation matrix to calculate the matching point pairs,which significantly enhances the likelihood of obtaining correct matches.The experimental results show that,compared with the point cloud registration driven by multi-feature extraction and matching matrix,the proposed method has higher registration accuracy and robustness to noise in partial point cloud registration.