Point cloud feature extraction method based on spatial attention mechanism

Abstract

The invention discloses a point cloud feature extraction method based on a spatial attention mechanism. The method comprises the following steps of: a, point cloud preprocessing: performing downsampling by utilizing a farthest point sampling algorithm; b, point cloud data set expansion: increasing the number of the point cloud training data sets by adding disturbance and random rotation, wherein point cloud data enhancement is not needed in the reasoning stage; c, point cloud feature acquisition: utilizing a double-flow network idea and one neural network branch to extract point cloud spatialfeatures, and using the other neural network branch to extract semantic features of point cloud overall information; and d, fusing semantic information of the two point cloud feature extraction networks layer by layer to obtain a feature vector of each point cloud. And e, after the point cloud features are obtained, performing further feature extraction on each point by utilizing a local space attention module, and finally obtaining a final classification or segmentation result through a full connection network. According to the method, a more accurate result can be obtained in a point cloud classification and segmentation task by using a double-flow network and a local space attention mechanism.

Description

technical field [0001] The invention belongs to the technical field of point cloud processing, and in particular relates to a point cloud feature extraction method based on a spatial attention mechanism. Background technique [0002] Point cloud is one of the most important basic representations in 3D processing. Usually, there are two methods for extracting point cloud features: one is to voxelize point cloud data and use 3D convolution to extract point cloud data features. This step consumes a lot of computing resources and slows down reasoning. slowness etc. Another method is to use a multi-view method to extract multiple view information of a 3D object, and then use a 2D image method for feature extraction. This method can achieve better results in the classification task of simple scenes, but in It does not play a big role in real scenes, and it is difficult to apply to the segmentation task of point clouds. [0003] At the same time, with the development of TOF tech...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to solve the problem that the spatial information of the point cloud cannot be fully utilized to extract the semantic features of each point of the point cloud under the condition of low calculation amount in the known technology, and to provide a point cloud based on the spatial attention mechanism. cloud feature extraction method

Images