Saliency-Guided Feature Matching for Self-Driving Systems

This paper presents a novel method for increasing the reliability of feature matching based on visual saliency. The key idea of the proposed method is to suppress outliers frequently occurring in the background clutter by adopting visually attractable (i.e., salient) regions as a filtering weight, which are highly relevant to underlying structures (e.g., signs, buildings, etc.) in a given scene. This significantly improves the accuracy of feature matching while reducing the memory usage. Experimental results on the KlTTI dataset demonstrate that the proposed method is effective for increasing the accuracy of feature matching under diverse outdoor environments.