Reserch On 3D Point Cloud Scene Flow Estimation Based On Optimal Transport

With the development and widespread application of Li DAR sensors,the cost of collecting point cloud data has significantly decreased,and point clouds have become the mainstream data type in machine vision tasks.The scene flow of point clouds can describe the three-dimensional motion information of objects,and the task of scene flow estimation has also received increasing attention.However,existing scene flow estimation methods have a large search range and high computational cost for global matching of point clouds.When extracting point cloud features,existing methods also do not consider the correlation information between point clouds,resulting in some errors and abnormal matching,reducing the estimation accuracy of the network.To address the above issues,this paper proposes a scene flow estimation method based on optimal transport theory.The main work includes the following aspects:Firstly,a regional point cloud scene flow estimation method based on optimal transmission is proposed to address the issues of high computational complexity,lack of guaranteed accuracy and reliability due to the majority of existing scene flow estimation methods using global matching.To solve the problem of point cloud feature extraction,a point cloud data feature extraction module based on Point Net++is designed;To solve the problem of noise caused by the large search range of point clouds in global matching,a regional scene flow estimation module is designed that divides regions based on key points;To solve the problem of significant differences in scene flows between neighboring points in region estimation methods,a scene flow optimization module based on point cloud neighborhood consistency was designed;Experimental research was conducted on the classic Flying Things3 D and KITTI datasets to verify the effectiveness of this method in large-scale point cloud scene flow estimation.Secondly,this paper proposes a new FPFH-Transformer structure that combines traditional point cloud feature fast point feature histograms with the Transformer structure to address the lack of attention to the internal geometric structural features of point clouds in point cloud applications,as well as the interpretability issues of models in deep learning.This structure does not lose the original geometric features of point clouds when learning internal and inter point cloud correlation features,further enhance the ability to extract point cloud features.Finally,this paper proposes a point cloud scene flow estimation method that integrates the FPFH-Transformer structure to address the issue of loss of correlation information between point clouds when extracting point cloud features in existing scene flow estimation methods.Through this module,the association information between points within the key point cloud and between point clouds is learned,improving the accuracy of key point matching;In order to address the dependency of scene flow estimates for each point in the region on a single key point in the region estimation method,a point to multi block scene flow estimation method was designed;A comprehensive loss function that can comprehensively measure the difference between the estimated value and the real value of the scene flow is proposed.Experiments on Flying Things3 D and KITTI datasets show that the method in this chapter can improve the matching accuracy between key point clouds,reduce outlier,and improve the performance of the entire scene flow estimation network.