Semantic Segmentation Of 3D Point Clouds Based On Deep Learning

In computer vision,3d semantic segmentation,as an important basis for 3d scene understanding,has been widely used in many 3d perception fields,such as: laser radar vehicle segmentation in unmanned driving,object shape segmentation in augmented reality and scene segmentation in robot obstacle avoidance.With the wide application of deep learning in the field of computer vision,deep learning on 3D point clouds has attracted more attention.PointNet is an advanced method on 3D point clouds.However,on 3D deep learning,the PointNet does not make great use of the local information between the point clouds,resulting in fuzzy segmentation and low precision.Aiming at the above limitations,this paper proposes a point cloud semantic segmentation algorithm based on KNN algorithm,aiming at improving the segmentation accuracy of point cloud.Then,based on this algorithm,aiming to improve the efficiency of deep learning,a semantic segmentation method of 3d point cloud based on super-voxel feature learning is proposed.The specific work of this paper is as follows:(1)We propose a 3d semantic segmentation method based on KNN,which increases the PointNet's abilities in extracting local features,and effectively improves the accuracy of semantic segmentation.The algorithm is divided into three parts: firstly,we implements the sub-module of point feature extraction based on PointNet,which is used to extract more powerful high-dimensional features of the original point cloud;secondly,the local feature extraction sub-module is established in the high-dimensional feature space,which uses KNN algorithm to search the K neighborhood of each query point to extract local features.Thirdly,the point cloud is semantically segmented by the global and local features.Experiments show that extracting local features between point clouds can improve the accuracy of semantic segmentation.Compared with PointNet,the overall accuracy of the algorithm is increased by 5.8%,and the mIoU is increased by 7.1%.(2)Based on the idea of data sparsity,we propose a fast 3D point cloud semantic segmentation method based on super-voxel feature learning,which accelerates the process of deep learning training and testing while ensuring segmentation accuracy.It is mainly divided into three steps: Firstly,the dense point cloud is segmented by super-voxel to generate sparse point cloud,and the normal and FPFH features are computed as learning features of the sparse point cloud.Secondly,the semantic segmentation network is trained based on sparse data to speed up the training process of the network.Finally,the rough segmentation results of semantic segmentation network are refined by using the full connection condition random field.Compared with PointNet,the proposed algorithm is 3.6 times faster on the overall time,5 times faster on training time,and 2 times faster on the test time.And in terms of segmentation accuracy,the overall accuracy oAcc is increased by 0.8%,and the mIoU is increased by 2.7%.Aiming at the semantic segmentation method based on KNN and semantic segmentation method based on super-voxelIn this paper,two models are tested and analyzed in detail based on the S3 DIS datasets.Compared with PointNet,the experimental results show that the two algorithms in this paper exceed the PointNet model in segmentation accuracy,which verifies the feasibility and effectiveness of the algorithm.