Complex scene 3D point cloud semantic segmentation method based on convolutional neural network

Abstract

The invention discloses a complex scene 3D point cloud semantic segmentation method based on a convolutional neural network, and the method comprises the steps: carrying out down-sampling of an obtained original point cloud, and carrying out feature extraction of a sampling point cloud obtained through sampling through a center self-attention mechanism and a neighborhood self-attention mechanism; splicing the extracted point cloud spatial position features and the obtained point cloud data attribute features, and obtaining a global feature vector through difference pooling processing under the attention mechanism; and cascading the up-sampling result of each layer and the corresponding global feature vector by adopting a jump connection mode, finally generating a point cloud segmentation neural network model through processing of a full connection layer, training and predicting the point cloud segmentation neural network model by utilizing multiple groups of obtained point cloud data sets, and finally completing a semantic segmentation task. Experimental results prove that the network model has higher generalization performance and good application value.

Description

Technical field[0001]The present invention relates to the field of computer visual technology, and more particularly to a complicated scene 3D point cloud language segmentation method based on convolutional neural network.Background technique[0002]In recent years, with the continuous development of laser radar equipment, RGB-D cameras, mature, 3D point cloud data quality, acquisition efficiency and cost performance continue to increase. As one of the computer visual long-term research topics, semantic segmentation is designed to use a computer to classify the scene, divide the scene into a number of regions with a specific semantic class, is the basis of many visual tasks 3D scenario understanding and analysis. It is schematically divided into two categories: Direct point cloud language segmentation and indirect point cloud language segmentation. In order to better learn the fine grain local characteristics of point cloud, some researchers attempt to introduce self-focus mechanisms ...

AI Technical Summary

Problems solved by technology

In order to better learn the fine-grained local features of point clouds, some researchers try to introduce the self-attention mechanism of adaptively screening local features into the network model, and further improve the semantic segmentation accuracy by ignoring irrelevant information and focusing on key information, such as self-attention The mechanism introduces GCNN to build the network GAPNet, and combines the self-attention mechanism and the cyclic neural network coding RNN to propose a convolutional neural network based on contextual attention. Combining the self-attention mechanism and random sampling algorithm, a large-scale 3D point cloud is designed. Lightweight point cloud semantic segmentation network RandLA-Net, but the self-attention mechanism in these networks focuses on learning the local structural features between the center of the sampling point and its neighborhood points, often ignoring the learning of the mutual structural relationship between the neighborhood points, and at the same time There is also little consideration of the role of the self-attention mechanism in the process of feature information network transmission. Semantic classes that are extremely similar in overall geometry and slightly different in local detail structures cannot be effectively distinguished, and there are varying degrees of under-segmentation problems; in addition, Although the GCNN-based point cloud semantic segmentation network has excellent performance, large-scale point cloud processing is a potential problem because the number of GCNN nodes is related to the number of point cloud points and the network structure is relatively fixed.

Images