Semantic Scene Segmentation of Unordered Point Clouds on Autonomous Robots

This paper describes a 3D semantic scene segmentation with convolutional neural networks for unordered point clouds of autonomous robots. Euclidean coordinates and RGB color spaces are used as well as multi-scaling layers. An outlier removal is designed to optimize the classification rate. We tested our system on real scenes using an RGB-D camera installed on a mobile robot. Additionally, we did comparison experiments on three different scene benchmarks. Compared to state-of-the-art point cloud semantic scene segmentation networks, our network produces better quality of segmentation results and achieves higher training and testing accuracies, as well as average intersection over union (IoU) and overall accuracy.