Deep Learning Method For Pose Estimation Of 3D Point Cloud With Multi-Scale Features

Three-dimensional point cloud is an important type of visual data,and tremendous application of point clouds have been developed in robotic,reverse engineering,medical imaging,self-driving car and other fields.The pose estimation of 3D object is one of the most important objectives of point cloud analysis.In the process of object recognition and pose estimation,point clouds captured by three-dimensional sensors will produce partial missing due to the occlusion and clutter in the scene,which makes it difficult for point clouds analysis and recognition.On the other hand,the unorganized data format of point clouds also challenges the application of deep learning for point clouds.Considering the above problems,this paper studies and designs a deep learning method for pose estimation of three dimensional point clouds with multi-scale features,by means of deep learning,data augmentation and multi-scale features extraction.Firstly,an interest point detection algorithm based on region growing clustering is proposed,and the algorithm is verified by the robustness and the descriptiveness experiments.Then the multi-scale method is investigated to compute the features of neighborhoods of point cloud interest points.Secondly,because of the scarcity of real annotated data,the training data for deep learning need to be enhanced.Therefore,a grid projection method is proposed to sample and obtain the point clouds from multiple perspectives,which can be used to simulate the data acquisition process of real 3D sensors.In terms of the network architecture,the spatial transformer network is adopted for random transformation parameter acquisition,and the parameters are transformed as the results of pose estimation.In order to obtain features of multi-scale local area independent of point order of point cloud,the symmetric function is utilized for pointwise features aggregation,and the multi-scale features are concatenated with the global features for both point cloud classification and pose estimation.The performance of the proposed interest point detection algorithm,the statistical performance of multi-scale features,the loss convergence and the visualization of the deep learning network are analyzed through experiments.The results show that the proposed interest point detection algorithm,deep learning network and multi-scale features are effective.Finally,the experiment of robot grasping in cluttered scene is conducted.After achieving the goals of point clouds segmentation,pose estimation,fine registration and grasping of particular objects,the effectiveness of the proposed method is verified.