Research On Feature Based Representation For Point Cloud Object Recognition

With the increasing demand of industry,the wide use of 3-D point cloud sensors and the increasing availability of point cloud data,the need for automatic processing of massive point cloud data is becoming increasingly urgent.It is of great theoretical and practical significance to study object recognition based on point cloud data.Feature extraction is the key to the question of object recognition.Considering that the traditional recognition method has a relatively complete theoretical basics,and the current deep learning based methods becoming increasingly popular among point cloud object recognition,this paper studies the problem of point cloud object recognition from two aspects: traditional based method and deep learning based method.1?A review of point cloud object recognition methods and pipeline is presented.Six widely used or newest point cloud descriptors are introduced,and the recognition performance of these six descriptors is compared in several aspects on the public datasets.2?The possible occlusions and inaccurate segmentation that occur in the scene will cause partial missing of the object,which will reduce the recognition accuracy of the global point cloud descriptors,and it is hard to achieve real time results for the recognition based on local descriptors.This paper propose a point cloud recognition method based on simulated occlusion,by the combination of the global and local surfaces,the proposed method improves the recognition accuracy effectively.This method is mainly implemented during offline stage and will not affect the time consumption of the online recognition stage.First,given the CAD models of the objects to be recognized,and the partial point cloud under different viewpoints of these models are obtained.Then,multiple direction vectors are obtained for each of the partial point cloud,and the point cloud is segmented into multiple subparts from each direction.Finally,multiple subparts are obtained for each of the CAD model,with some subparts having significant overlap.The ICP algorithm is used to remove the redundant surfaces,which is effective to reduce the size of the model library while maintaining the recognition rate.Experimental results on public datasets show that the proposed method promotes the recognition rate significantly.3?The 3-D points of the objects have different contributions to the final recognition result,and the 3-D points have different discriminating capabilities.Some 3-D points on the edge of the objects should be more helpful for recognition.However,some point cloud processing networks use the farthest point sampling method to obtain the key points.This paper propose to add an attention module to give a larger weight to the edge points in these key points.First,several attentional networks for image classification are introduced.Then,a novel attentional module was proposed and add to the existing point cloud processing network.Experimental results on public datasets show that the proposed method promotes the recognition rate effectively.