Design And Implementation Of Point Cloud Semantic Segmentation Network Based On Octave Search

In recent years,with the continuous development of three-dimensional point cloud acquisition equipment,point cloud processing has become a research hotspot.As one of the important tasks of point cloud processing,point cloud semantic segmentation has gradually proved to be the most effective means to understand point cloud from the perspective of fine granularity.It has been widely used in practical scenes such as autonomous driving,intelligent robots and indoor scene analysis.However,due to the characteristics of point cloud and the influence of acquisition equipment,it is still a challenge for current point cloud research to design a network that directly processes the semantic segmentation of point cloud data.This paper mainly studies the semantic segmentation task of point cloud and proposes a point cloud semantic segmentation network based on octave search.The main research work of this paper is as follows:(1)In this paper,the local feature extraction process of point cloud is divided into five steps:key point sampling,local area construction,feature fusion,learning transformation matrix and feature extraction.In the local area construction,this paper proposes a new method of adjacent point sampling,namely octave search.The octave search method divides the spatial region around the key point into eight sub-spaces,and builds local areas based on the neighboring points in each sub-space.In order to accelerate the convergence of the matrix parameters,the transformation matrix is optimized,which makes the transformation matrix more reasonable and effective in dealing with the ordering of point cloud features.(2)In the aspect of semantic segmentation network construction,this paper follows the Encoder-Decoder architecture,and uses the skip-link structure in the Decoder stage to integrate the feature information between the shallow and deep network,thus ensuring that each point in the output results can be allocated a reasonable semantic tag.(3)In order to verify the performance of the proposed network,this paper has trained and tested in different data sets.The experimental results show that the segmentation network in this paper can obtain relatively ideal results on each data set.In addition,in the network training,this article uses the multi-process and multi-GPU technology,significantly improving the training speed of the network.