| As the basis and key of realizing 3d scene understanding,semantic segmentation of 3D point cloud is widely used in the fields of autonomous mobile robots and autonomous driving.Due to the small targets in the outdoor point cloud,such as pedestrians and cyclists,cannot be ignored for automatic driving,it is necessary to achieve precise semantic segmentation of small targets.Traditional point cloud segmentation algorithm is not suitable for large outdoor scene,the existing deep learning semantic segmentation algorithms are not accurate enough to segment small objects.Therefore,this paper focuses on data enhancement,loss function design,local and multi-scale context feature extraction of small targets in semantic segmentation of outdoor point clouds,including the following contents:1.Aiming at the semantic segmentation problem of small targets in outdoor point cloud,this paper proposes hierarchical random sampling to enhance the retention of more points and designs a small target segmentation algorithm based on context attention.The algorithm mainly consists of bidirectional context attention fusion(BCAF)module and context code-channel self-attention(CE-CSA)module.BCAF module fuses adjacent scale features through dual-channel bidirectional attention,effectively fuses shallow features and high-level semantic features.CE-CSA pays attention to relevant channel features selectively by context encoding multi-scale features and using channel self-attention mechanism.Experiments on Semantic KITTI and Semantic 3D data sets prove that the proposed network has high recognition accuracy for small targets and can accurately identify some object edge points2.Aiming at the problem of small target feature extraction in outdoor point cloud,since the local neighborhood feature extraction of PointNet++ is not sufficient,this paper proposes a new local feature encoding module to effectively encode the geometric location information.This module combines it with the attribute information coding feature to effectively enhance the local feature extraction.In addition,weight parameters and measurement indicators are introduced into the loss function to make up for the class imbalance problem.Experimental results on Semantic 3D data set show that the proposed feature encoding module is beneficial to the segmentation of small targets,and the improvement of loss function is beneficial to the improvement of segmentation accuracy.3.Aiming at the problem of missing small target in outdoor point cloud,this paper designs a small target semantic segmentation algorithm based on asymptotic optimization.Since small targets account for a small proportion in the cloud of scenic spots,this paper optimized the segmentation result by combining the two branches of thickness and fine.The coarse segmentation branch adopts the CAF-Net proposed in the first point for segmentation.The fine segmentation branch first detects the small target region of interest,and then uses the improved PointNet++ to perform fine segmentation on the local point cloud cluster of interest.Experimental results on Semantic KITTI show that the proposed method has better segmentation performance for small targets. |
