Research On Point Cloud Registration Algorithm Based On Deep Learning

Point cloud registration technology has become a research hotspot due to the need for fast mapping in computer vision fields such as auto-driving,medical imaging,artificial intelligence,high-precision maps and so on.The existing traditional point cloud registration methods can achieve a high registration accuracy,but they require a high initial posture,have limitations such as manual descriptors,iterative optimization and so on,and have low computational complexity,high efficiency,and can not meet the needs of dynamic learning features.In recent years,point cloud registration using in-depth learning method has become the mainstream trend,but the existing in-depth learning methods are based on high overlap of interference-free data training.There are problems such as low registration accuracy,low robustness and low generalization ability under noise,low overlap,etc.Therefore,to solve the two problems of low registration accuracy and poor generalization ability of existing deep point cloud registration methods under noise and low overlap conditions,this paper presents an Offset Cross Attention Dynamic Graph Convolutional Neural Network(OCADGCNN),which inserts an offset attention module into the dynamic convolution neural network to extract global feature vectors.Make full use of local structure information and spatial semantic information of point cloud to reduce information loss;Residual connections are added to improve network performance,and aggregation pooling modules are designed to further optimize feature extraction.The interactive attention module is used to exchange information between global features,enhance related information,and suppress the interference of non-overlapping area information.The results show that OCADGCNN has the lowest registration error compared with ICP,Point Net LK,PCR Net,OMNet and DOPNet in both noise-free and noise-free situations.Compared with DOPNet,the mean square error of rotation and shift in noise-free situations are reduced by 0.06% and 3.70%,and 0.75% and 3.33% in noise-free cases.The OCADGCNN model is robust to low overlap disturbances and its rationality is validated by experiments with reduced overlap and ablation.Secondly,based on OCADGCNN,an offset cross-focus end-to-end cloud registration algorithm(T-OCADGCNN)is proposed which fuses Transformer networks to handle complex semantic information interaction problems.Find the spatial relationship between different features using the multi-head attention mechanism.The key point extraction module is designed to ensure the accuracy and stability of registration.The corresponding point generation module is applied to adapt to the sparse,occlusion and other disturbing factors in local point cloud registration to ensure the correct mapping relationship is established.Compared with OCADGCNN,the mean square errors of rotation and translation are reduced by5.37% and 12.50% respectively in the case of noise,1.51% and 17.24% in the case of noise.The anti-jamming ability of T-OCADGCNN to noise and its processing ability on complex asymmetric models are verified,and the generalization and universality of the models are strong.The anti-interference ability of the model and the validity of the model for low overlap are verified by the overlap reduction test and ablation experiments.