Vision-based traversability estimation in field environments

Robust obstacle detection and traversability estimation remain a challenges for mobile robots traversing outdoor field environments. Illumination and environmental variances limit the applicability of appearance cues, while vegetation limits structure cues. Systems that combine multiple cues can potentially overcome deficiencies in individual cues. A key challenge in designing multi-sensor systems is to automatically and appropriately combine these cues in an unsupervised manner. This thesis presents methods for online obstacle detection and traversability estimation in field environments which continuously learn online about environmental and illumination conditions, and can operate in the presence of significant vegetation. The results demonstrate these methods in online field experiments and show that they give competitive performance without the requirement of pre-training or environment-specific tuning.