Transformer-Based Traversability Analysis for Autonomous Navigation in Outdoor Environments With Water Hazard

Traversability analysis is of vital importance for autonomous vehicle navigation in outdoor environments. This paper focuses on the navigation algorithms in outdoor environments with water hazards, where the water areas can neither simply be neglected nor be deemed as non-traversable for safe and efficient navigation. To this end, we present a tightly-integrated navigation algorithms framework, which includes an environmental traversability analysis model and a control signal acquisition model. The traversability analysis model includes the SwinT-CGRU semantic segmentation model and the ConvCRFs water analysis model proposed in this paper. SwinT-CGRU can efficiently extract the feature from the input successive images, providing initial analysis of environmental traversability. To further analyze the water region, we employ a ConvCRFs, which determines the water traversability in an effective way. The results of the water traversability analysis are convincing for reliable vehicle navigation in outdoor scenarios. Based on the traversability analysis results, we propose a path planning approach by adapting the artificial potential field method to the image space. The control signal for steering and driving the vehicle can be acquired through an established mapping from the image space to the vehicle space. Experiments demonstrate that our segmentation model improves mIoU by 2.09-18.21% on the RELLIS-3D dataset. Furthermore, our navigation algorithm exhibits real-time performance and stability, ensuring the safety of vehicles in outdoor environments.