Robust Depth-Aided Segmentation for Drivable Region Detection in Challenging Environments

Vipul Vijay Ramtekkar, Liza Dahiya, Nikunj Shah, Kenji Nishimiya, Takahiro Kuroki, Chaehyeon Song, Ayoung Kim, Myung-Hwan Jeon

Keywords: Drivable regions, RGB-D data, Semantic segmentation, SegFormer architecture, Depth-based refinement, Edge computing devices, Complex terrains, Real-time detection

TL;DR: This paper proposes a method using RGB-D data to detect drivable areas in challenging terrains, enhancing accuracy with SegFormer architecture and depth-based refinement for real-world reliability, achieving precise results on edge devices.

Abstract:

This paper proposes a method for detecting
drivable regions in challenging terrains using RGB-D data.
By integrating depth information with semantic segmentation,
our approach significantly improves detection accuracy across
diverse landscapes. Leveraging the SegFormer architecture,
we effectively distinguish drivable from non-drivable areas.
Additionally, we introduce a depth-based refinement mechanism
to ensure reliable performance in real-world scenarios. Extensive
evaluation in both off-road and on-road environments confirms
the effectiveness of our approach. Using the SA-1B dataset
with grounded SAM, our method achieves precise delineation
of road classes during training. Overall, this work advances
autonomous navigation systems by providing a comprehensive
solution for drivable region detection in complex terrains in
real time, even on edge computing devices.

Submission Number: 11