A RCDNet for outdoor unmanned sweeping vehicles based on graph convolutional network

• Published in: 2021 33rd Chinese Control and Decision Conference (CCDC) pp. 3133 - 3138
• Main Authors: Yu, Dianyong, Bian, Saiyingnan, Hu, Changan, Lou, Jieming
• Format: Conference Proceeding
• Language: English
• Published: IEEE 22.05.2021
• Subjects: Adaptation models; Convolution; Data models; Deep learning; Feature extraction; Graph convolution Point cloud; Local features; RCDNet; Roads; Urban areas
• ISSN: 1948-9447
• DOI: 10.1109/CCDC52312.2021.9601845

This paper proposes a graph convolutional network for road curb point detection (RCDNet) which converts road curb point detection into a point cloud segmentation problem. The input of the network comes from the point cloud below the segmented road edge. First, the point cloud enters a permutation network to ensure that the spatial permutability of the point cloud remains unchanged. Then a neighbor number k is given, and k neighbor points of each point are calculated through the xyz attribute of the point cloud. These k neighbor points used as the local features of the point are fused with the xyz attributes of the point to construct the edge feature. Then convolving the edge features to extract higher-level point cloud spatial information. Through this method of fusion of local features and global features, more local information of the point cloud can be obtained, so that the network can segment the three-dimensional point cloud with higher accuracy.