From one to many: Unsupervised traversable area segmentation in off-road environment

Abstract

Traversable area segmentation is important for safe navigation of mobile robot in outdoor environment. To address this problem, we propose a unified framework to register data across sessions, on which an unsupervised method is presented for traversable area segmentation intended for unstructured environments. With data collected on a vehicle equipped with camera and laser, the proposed method can generate massive label images for traversable and obstacle area without any human intervention, which are fed as training samples of a pixel-wise semantic neural network. In deployment, only a monocular camera is needed to work with the trained network, without structured assumption of the road such as lanes and traffic signs. The proposed method is validated on 4 datasets to demonstrate performance on traversable area segmentation. Moreover, it is shown that our method can be generalized to varied appearance at different location and time with distinct sensors.