The Synthetic Off-road Trail Dataset for Unmanned Motorcycle

The Unmanned Motorcycle (UM) is a robotic system aiming to autonomously drive in off-road trail environments, keep the tracked routes under surveillance, and provide help within its ability. One core component of the UM is its vision module that utilizes deep learning techniques to get pixel-accurate trail understanding. By learning valuable features from the off-road trail datasets, the UM is expected to understand traversable areas in a large variety of off-road trail scenes. However, currently, there is no publicly available dataset on the off-road trail scene dataset for UMs, which motivates us to conduct a comprehensive dataset generation and collection. In this paper, we build four virtual worlds to generate Synthetic Off-road Trail (SORT) dataset. The dataset seeks to positively influence the development of data-driven trail segmentation for UMs, which we hope other UM researchers will use the dataset and contribute to it. By this method, the dataset can be easily regenerated and tailored for other robot platforms. The whole workflow of tackling the scene parsing task of the UM is provided, and we hope our work can inspire the robot perception research community and help push the usage of realistic visual simulation to develop learningbased algorithms in their research fields. Multimedia material of our dataset is available at https://www.youtube.com/watch?v= 9biV7fKxKRo&list=PLxSoUS6AFfb -KlqzxpqlMfKoF9qur4h0p.