Research On Denoising Processing And Feature Description For 3D Point Cloud

In recent years,3D point cloud,a new primitive representation for the object,has been widely used in the research field of 3D object recognition due to its simplicity and expressive ability.However,because of the limitation of 3D acquiring devices,the captured 3D point cloud is inevitably contaminated by noise.It is,therefore,necessary to filter raw point clouds to produce high-quality and high-accuracy models laying the foundation for 3D object recognition.On the other hand,3D feature extraction is another important step in 3D object recognition,including hand-crafted features and deep learning-based features.The present hand-crafted features donot achieve ideal recognition performance overall.Although the rapid development of deep learning technology has brought new opportunities and breakthroughs,the characteristics of 3D point cloud data determine that it will be a challenging problem to adopt deep learning technology for 3D point cloud feature learning.Therefore,what strategies used and how to extract important features are still the most worthy questions.Based on this,this thesis pays attention to two main problems arising in 3D point cloud based object recognition,namely,denoising and feature description.The main contents are as follows:(1)For computational efficiency of 3D point cloud filtering algorithms,a type of guided 3D point cloud filtering approach,therefore,is proposed.The basic idea behind this proposal is that it takes the positions of points into account to derive a linear model concerning the guidance point cloud(input point cloud is used in this method)and the filtered point cloud.Then,an optimization strategy is applied to minimize the defined cost function to estimate the corresponding coefficients for the linear model.Experimental results show that the proposed algorithm can successfully remove the noise and is significantly superior to several state-of-the-art methods,particularly in terms of efficiency.And in the process of 3D point cloud based object recognition,this algorithm can help to improve the representative ability of feature descriptors,and recognition performance,more suitable for real-time 3D object recognition tasks.(2)For the process of different levels of noise,especially high-level noise,a simple but effective 3D point cloud filter is introduced based on the idea of iteration and normals information.In this proposed framework,the ameliorated principal component analysis algorithm is exploited to estimate surface normal for each point.Afterward,a new joint bilateral filter based on spatial information and normal information is defined to apply to the normal filed,which will be performed iteratively.It is important to note that in the process of iteration,the normal field obtained from the previous iteration is used as the guidance for the current iteration.Finally,the improved point updating strategy is performed to adjust point positions to match the filtered normals.Experimental results indicate the proposed algorithm's simplicity,effectiveness,and ability to deal with high-level noise.And its application in 3D point cloud based object recognition shows that this algorithm lays the foundation for the improvement of the ability of feature extraction and recognition,which is suitable for object recognition tasks requiring high-quality point cloud.(3)For the effect of the hand-crafted feature on the performance of 3D point cloud based object recognition,based on analysis of the advantages of local-based and global-based point cloud descriptors,this thesis proposes the Local-to-Global histogram features(LGH for short),which includes the local viewpoint feature histogram and the local ensemble of shape function.And this kind of Local-to-Global histogram feature depends on the sub-point cloud determined by keypoint extracted using 3D keypoint detection algorithms and its neighborhood and global-based histogram features.Experimental results show that these algorithms have better performance than the selected methods as the comparison in terms of descriptiveness,robustness to noise,efficiency,combination with different 3D keypoint detection algorithms,and with different filters.The results also demonstrate that the proposed guided 3D point cloud filter and iterative guidance normal filter are conducive to the improvement of expressiveness of Local-to-Global histogram features.(4)For the problem brought by the unstructured and irregular property of point cloud to deep learning-based tasks,this thesis proposes a two-stage framework for point cloud feature learning.The first step performs voxelization operation on point cloud to form the regular voxel grid using the binary occupancy grid.Then,the voxelized data are fed into a new designed 3D dense-attention convolutional neural network framework(including channel-wise attention module based 3D convolutional neural network,namely 3DDACNN-CA and channel-spatial attention module based3 D convolutional neural network,called 3DDACNN-CSA).Since the network adopts a much deeper structure and integrates the attention mechanism,the learned features have more powerful and discriminative representation.Experimental results also show that the guided point cloud filtering and iterative guidance normal filter play positive roles in improving the expressive ability of deep learning features.