3D SEMANTIC PERCEPTION OF INDOOR ENVIRONMENTS

Abstract

This thesis is focused on studying the issue - 3D semantic perception of indoor environments. We start this study with a plane mapping system, which takes advantage of both 3D dense reconstruction and plane detection approaches. 3D semantic labeling is an interesting yet challenging problem for indoor scenes. To circumvent noisy 2D semantic segmentation results, we perform semantic labeling on RGB-D video sequences instead. Long-term spatio-temporal relationships among all observed 3D points can be generated by detecting spatial connected components. Our approach significantly improves semantic labeling accuracy and consistency. Point cloud is a simple and straight-forward representation of 3D structures. We propose the PointAtrousGraph - a graph-based architecture for hierarchical feature learning on point clouds. We design graph-based operation modules to propagate multi-scale point features. Experiments show that the proposed PointAtrousGraph achieves better performance compared to previous state-of-the-art methods in various applications.