Learning Off-Road Terrain Traversability with Self-Supervisions Only

Estimating the traversability of terrain should be reliable and accurate in diverse conditions for autonomous driving off-road environments. However, learning-based approaches often yield unreliable results when confronted with unfamiliar contexts, it is challenging to obtain manual annotations frequently new circumstances. In this paper, we introduce a method learning from images that utilizes only self-supervision no labels, enabling easily learn To end, first generate self-supervised labels past trajectories by labeling regions traversed vehicle as highly traversable. Using then train neural network identifies terrains are safe traverse an image using one-class classification algorithm. Additionally, supplement limitations incorporating methods visual representations. conduct comprehensive evaluation, collect data variety environments perceptual show our produces estimations various addition, experimental validate outperforms other estimation achieves comparable performances supervised trained on manually labeled data.