Registration,Recognition And Labeling In 3D Point Clouds

With the continuous development of sensor technology,the manners of human perceiving world has been developed from the traditional 2D sensing technology to the 3D sensing technology.Specially,with the development of 3D laser imaging technology,3D laser scaning system has been widely used to get 3D point clouds of objects or scenes.The number of 3D data is achieving a significant growth,making that 3D object modeling and scene modeling,3D object recognition and scene understanding has obtained great progress.However,point cloud registration,3D object recognition and 3D scene labeling are still important and challenging tasks in 3D computer vision.Motivated by these challenges,this thesis presents an extensive theoretic and technical investigation on point cloud registration,3D object recognition,and 3D scene semantic segmentation.For 3D point cloud registration,an efficient rotation estimation algorithm is proposed for point clouds of structured scenes.The registration consists of two parts: rotation estimation and translation estimation.For rotation estimation,a direction angle is defined for a point cloud based on their normals.Then,then the rotation matrix is obtained by comparing the difference between the distributions of direction angles.To conduct a full3 D registration,the translation parameters are estimated by a simplified Iterative Closest Point(ICP)algorithm.Compared with the feature based registration algorithms,the proposed algorithm can be used for the registration task of structured scene with less feature.Experimental results demonstrate that the proposed algorithm outperforms the state-ofthe-art approaches in terms of both accuracy,computational efficiency and robustness.For 3D object recognition,a 2D projective views based multi-view convolutional neural network boosted by view saliency is proposed for feature learning and recognition.Multi-view projection views are first obtained by rendering the 3D object with virtual cameras.Then,the discriminativeness of each view is investigated with view saliency using 2D Zernike Moments.Finally,the proposed view saliency is then used to boost a multi-view convolutional neural network for 3D object recognition.The proposed method can effectively reduce the information redundancy in multi-view feature fusion.Experimental results have shown that the performance of the proposed method has been significantly improved over the existing methods.For 3D point cloud labeling,deep multi-scale feature learning neural networks are proposed.On one side,the unordered point clouds are taken as input data and multi-scale feature learning is applied on the spherical neighborhood of each point.Specifically,the neighborhood is conducted using Euclidean distance between points.A local and global feature aggregation block is also introduced to the hidden layers in the network to improve its feature learning capability.On the other side,to improve the efficient of multi-scale feature learning,a point cloud and voxel model based neural network is proposed.The point cloud based local network can be used to extract the local information of each point.The voxel based network can be used to extract the spatial context information between each point.The combination of the two can effectively the feature learning capability.Furthermore,to improve the smoothness of point cloud labeling,a recursive convolutional neural network based fully connected conditional random field is used for later optimization.Experimental results have clearly demonstrated that the proposed networks achieve high labeling performance and computational efficiency.