Point cloud semantic segmentation method for color difference guided convolution

Abstract

The invention provides a color difference guided convolution point cloud semantic segmentation method. The method comprises the steps of converting colored point cloud RGB into HSV; for the point cloud under the current density, solving a k neighborhood to obtain a relative position y of a central point and a neighborhood point, and sending the relative position y to an MLP to obtain a feature FG;obtaining d1, d2 and d3 in three directions according to neighborhood different channel color moment sorting and y, and selecting corresponding features of nearby points; convolution is performed onthe selected features in three directions and color channels, so that features FC can be obtained; cascading the FG with the FC to obtain a global feature F; performing down-sampling, and repeating the steps 2-5; f and FC interpolations under adjacent densities are recovered, and a prediction result of each point is obtained; based on the cross entropy loss function, minimizing the loss function through gradient descent, and training neural network parameters; and after the parameters are trained, when a new to-be-segmented point is given, executing the steps 1-8 to obtain a segmentation result. It can be seen from experiment results that the method can obviously improve the point cloud semantic segmentation precision under various types, and is suitable for indoor and outdoor scenes.

Description

technical field [0001] The invention relates to the technical field of semantic segmentation of small-scale point cloud images, in particular to a method for point selection and convolution feature extraction guided by point cloud neighborhood color information in indoor and outdoor point cloud semantic segmentation. Background technique [0002] Three-dimensional data has a strong ability to reflect the real scene, and has been paid more and more attention by researchers. Point cloud is the main format of 3D data. Semantic segmentation of point cloud is a necessary work for scene understanding and the key to robot development, autonomous driving, virtual reality and remote sensing mapping. Inspired by the success of deep learning methods for 2D images and 1D text, many researchers have applied these techniques to analyze 3D point clouds. But point clouds are intractable as direct input because they are inherently unstructured and unordered. [0003] PointNet is a mileston...

AI Technical Summary

Problems solved by technology

It inputs the entire point set into a shared multi-layer perceptron (MLP) for convolution, and enables PointNet to process unordered points through symmetric pooling operations, but does not consider neighborhood information in PointNet

Images