Feature Matching under Region-Based Constraints for Robust Epipolar Geometry Estimation

Outlier-free inter-frame feature matches are important to accurate epipolar geometry estimation for many vision and robotics applications. We discover a set of high-level geometric and appearance constraints on low-level feature matches by exploiting reliable region matching results. A new outlier filtering scheme based on these constraints is proposed that can be combined with traditional robust statistical methods to identify outlier feature matches more reliably and efficiently. The proposed filtering scheme is tested in an real application of outdoor mobile robot navigation based on far-field scenes and of scenes that contain repeated structures.