2DLaserNet: A deep learning architecture on 2D laser scans for semantic classification of mobile robot locations

In this work, we deal with classification of mobile robot locations into semantic categories such as room, corridor, and doorway using 2D laser data. Previous studies were generally able to distinguish room and corridor classes; however, the classification of doorway locations has not been satisfactory. To increase the classification accuracy of doorway class, we proposed a new point-based deep learning architecture, namely 2DLaserNet. In contrast to the well-known point-based deep learning techniques, 2DLaserNet exploits the ordered relation between successive points in the point cloud generated from 2D laser readings. In this way, 2DLaserNet is able to learn the geometric characteristics of laser scans corresponding to room, corridor, and doorway classes. We used the publicly available Freiburg 79 dataset to validate the effectiveness of the proposed approach, especially for the doorway class. Besides, we incorporated synthetic data to account for the intra-class variety for doorway locations. We also conducted experiments on the Freiburg 52 test dataset to examine the generalization ability of the proposed architecture trained with the Freiburg 79 dataset. We observed that 2DLaserNet outperforms state-of-the-art methods and well-known point-based deep learning techniques for doorway class.