Traversability Analysis for Unmanned Ground Vehicles Based on Multi-modal Information Fusion

How to efficiently and safely navigate unmanned ground vehicles in unstructured off-road environments still faces several challenges, such as various terrains and highly dynamic scenarios. Traversability analysis presents tremendous potential in addressing these problems by perceiving surroundings and building travesability maps. In this paper, a novel traversability analysis algorithm based on multi-modal information fusion has been proposed. This algorithm comprehensively leverages data from both LiDAR and camera to understand the characteristics of the environment. Point clouds are utilized to extract geometric features such as flat surfaces, slopes and depressions, and semantic segmentation of images enables the identification of various terrains as well as dynamic objects. These two types of information will be effectively fused and integrated to generate a traversability map on the fly, which is dynamically insensitive. To verify the performance and effectiveness, the algorithm has been deployed on an Unmanned Ground Vehicle(UGV) where MMDeploy Toolbox has been used to accelerate the inference speed of the segmentation model and maintain the input stream frequency of mapping at around 10 Hz. Extensive experiments have been conducted within campus and the algorithm demonstrates robustness in such environment with diverse terrains and dynamic objects.