Study Of Point Cloud Semantic Segmentation Of Outdoor Large Scene Algorithm Based On Deep Learning

With the development of laser scanning technology,it is more and more convenient to obtain dense 3D point cloud data in outdoor scenes.As one of the key technologies of automatic driving,point cloud semantic segmentation has become the research emphasis in navigation and positioning,computer vision,pattern recognition and other fields.Therefore,it is of great significance to study semantic segmentation under large outdoor scenes.At present,deep learning has been widely applied in 2D image field.However,some characteristics of point cloud,like disorder and non-structure,have bring some challenges to point cloud semantic segmentation based on deep learning.Therefore,this paper will lucubrate the algorithm of point cloud semantic segmentation in outdoor scenes adopting the methods of deep learning.Firstly,This paper proposes an outdoor point cloud semantic segmentation algorithm based on improved PointNet++,explores the flow of PointNet++processing point cloud from outdoor large scenes,so that it can be successfully applied to point cloud of outdoor scenes.Since PointNet++ did not consider the global information of input point cloud,this paper extends the SENet module to 3D space to embed the global information of point cloud through learning the relationship between 3D point cloud feature channels,so that the extracted features could be more discriminating.The experimental results show that the improved network achieves higher segmentation accuracy.Then,combined with the scale space theory,this paper proposes a point cloud semantic segmentation network based on multi-level scale space.The multi-level scale space of point cloud includes both multi-scale density and multi-scale space.This paper takes multi-scale point cloud as the network input,extracts point cloud features by PointNet++and then aggregates point cloud features of different scales.Besides,three models are designed to test segmentation effect,which is based on multi-scale density,multi-scale space and multi-level scale space respectively.The results of experiments show that the method of point cloud multilevel scale space can effectively segmentate point cloud of outdoor scenes and gain obviously better performance than native PointNet++.Finally,a point cloud semantic segmentation network based on spatial neighbor information is proposed,which makes use of the advantage of Gate Recurrent Unit(GRU)to learn long-distance dependence.Firstly,the whole 3D cloud space is split into uniformly-spaced blocks on the ground,the PointNet++module and max pooling operations are respectively used to get the local features and the global features of each point cloud.Then the GRU is used to update the sequence of each block features.Finally,the global features,local features and the updated features are fused.In addition,this paper also tries to use multi-level PointSIFT module to extract features.The experimental results show that the proposed point cloud semantic segmentation network based on spatial neighbor information can achieve better performance in point cloud semantic segmentation of outdoor scenes.