Pavement scene interpretation and obstacle detection by large margin image labeling

This paper presents a novel discriminative approach for pave-ment scene understanding and obstacle detection in real-world images. It overcomes the heavy constraints in previous systems such as a simple background, a specic obstacle, etc. The approach we exploited extends the bundle method to incorporate pairwise correlations among neighboring pixels, and adopts graph-cuts as the inference engine to attain the approximation efficiently. A set of robust features on both local and multi-scale level is also introduced that captures the general statistical properties of pavements and obstacles. The proposed approach is validated on real-world image database, and outperforms the current state-of-the-art visioned-based methods