Spatially Coherent Costmap: A Weakly Supervised Pipeline for Outdoor Traversability
Year: 2024
Type: conference-paper
Abstract: Terrain perception is crucial for autonomous navigation, especially in off-road environments where terrain variability can pose challenges. Beyond simple obstacle avoidance, autonomous navigation can benefit from finer terrain traversability analysis. We introduce a novel method for creating terrain traversability maps, bringing benefits in terms of navigation safety and efficiency. The challenge of defining and manually annotating terrain traversability advocates for the use of limited annotations. Our method introduces a new weakly supervised approach with Siamese Networks to estimate trajectory cost from IMU and a generic Segmentation model to extend this annotation to full images. This methodology overcomes data sparsity and the limitations of existing formulas that struggle with simple terrain discrimination (like gravel vs road) while keeping very low annotation effort.
Source: HAL
Institution: Unité d'Informatique et d'Ingénierie des Systèmes
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7207092740
