Research On Point Cloud Completion Method Combined With Key Points Sampling And Multi-scale Feature Fusion

As an intuitive representation of 3D object acquired by lidar and other special equipment,point cloud is widely used in many fields,such as virtual reality,autonomous driving,engineering surveying and mapping.Compared with traditional RGB image,point cloud can capture an object precisely and describe it with unsorted coordinate sequences.Nonetheless,the scanned data are often incomplete due to the underperformance of the devices and poor surroundings.This problem greatly affects the development of subsequent tasks.Therefore,this thesis takes partial point cloud data as the main research target and proposes a method that combines key points sampling and multi-scale feature fusion to achieve high-quality point cloud shape completion.Firstly,because of the large scale of the point cloud collection,it is necessary to reduce the data size before training the completion network.Aiming at the problem that the existing point cloud down-sampling algorithms have poor ability to represent key points,this thesis proposes a key points sampling algorithm based on cosine similarity.This method uses the similarity measure between vectors to evaluate the geometric structure complexity of the local area in the point cloud.So as to flexibly select sampling points.Secondly,in order to make the completion network has a better perception of object details,multi-branch networks are often used to extract the features of point clouds with different resolutions.Aiming at the problem that existing multi-scale completion networks have weak correlation in between-scales features,and redundant information appears in within-scale features,this thesis proposes a multi-scale feature fusion method named cross cascade module to realize the information fusion between different scales.This module uses cross modeling and leap type convolution to enrich the expressive ability of local features and broaden the receptive field with an unchanged kernel size.Finally,corresponding to the problem that the global completion strategy will change the overall data distribution and cause a distorted point cloud,this thesis proposes partial shape mask method and hierarchical feature adaptive offset method to make the network focus on the dense reconstruction of missing structures guided by fine-grained component category,so as to reduce the reconstruction error.In addition,this thesis also combines cross stage partial structure with cross cascade module to improve network learning ability and reduce computational consumption by splicing and reusing separated features.Results on the Shape Net-part dataset show that the proposed network can effectively reduce the reconstruction error and improve the quality of point cloud shape completion.