Point cloud identification and segmentation method based on Bayesian neural network

Abstract

The invention relates to a point cloud recognition and segmentation method based on a Bayesian neural network, and the method builds the Bayesian neural network, and comprises three parts: feature extraction, recognition and segmentation. Wherein the feature extraction part comprises three Bayesian convolution layers, and an activation layer is connected behind each convolution layer. The networkstructure of the identification part is a three-layer full-connection layer, and the number of nodes of the final output layer is the same as the number of categories. Wherein the network structure ofthe segmentation part is divided into three Bayesian convolution layers, an active layer is connected behind each convolution layer, the number of nodes of the final output layer is equal to the number of points contained in the point cloud, and each node outputs a vector of which the length is equal to the number of categories. The point cloud data firstly pass through the feature extraction part to obtain feature values, and then the feature values are respectively input into the recognition part and the segmentation part to obtain recognition and segmentation results.

Description

technical field [0001] The invention relates to the application of deep neural network and Bayesian probability in computer vision, in particular to a point cloud recognition and segmentation method based on Bayesian neural network. Background technique [0002] Deep neural network is a very popular research direction in the field of artificial intelligence in recent years, and breakthroughs have been made in image processing, target recognition, and speech recognition. With the development of three-dimensional sensors such as lidar and structured light, deep neural networks have gradually been applied to the recognition and segmentation of point cloud data. Use the multi-layer perceptron with shared weights to extract the depth features of point coordinates and related additional information, and use the method of pooling and feature splicing to extract the local correlation features between points, and finally use the fully connected layer to realize the recognition of poi...

AI Technical Summary

Problems solved by technology

[0003] Dai Lu, Wang Junliang et al. ("Non-equivalent point cloud segmentation method based on convolutional neural network", Journal of Donghua University (Natural Science Edition), 2019, 45(6):862—868) Aiming at the non-equivalence in point cloud segmentation, a CNN-based point cloud segmentation neural network is proposed to solve the variable amount and sequence of point cloud data in the design network random sampling layer and maximum pooling layer On the basis of , the distance matrix after the penalty function is introduced to weight the classification errors of each point to optimize the model training method. However, due to the inaccuracy of the artificially designed penalty function design, this method cannot guarantee the optimal model training effect.

Images