Robust Invariant Features for Object Recognition, Pose Estimation and Topological Navigation

Robust Invariant Features for Object Recognition, Pose Estimation and Topological Navigation

  We present a new robust image feature detector for the object recognition and vision based mobile robot navigation. The proposed algorithm extracts highly robust and repeatable features based on the key idea of tracking and grouping multi-scale interest points and selecting a unique representative structure with the strongest response in both spatial and scale domains. Weighted Zernike moments are used as the local descriptor for feature representation. The experimental results and performance evaluation show that our feature detector has high repeatability and invariance to large scale, viewpoint and illumination changes. The efficiency and usefulness of the proposed feature detection method are also confirmed by the excellent performance on object recognition and mobile robot indoor navigation.