A New Robotic Context-Based Object Recognition Algorithm for Humanoid Robots

Abstract

This paper proposes a new object recognition algorithm using robotic context information for humanoid robots. For more robust object recognition for less textured objects, we combine shape-based interest points and local appearance-based descriptors computed in a neighborhood around each detected interest point. The combination is used as a basic feature for matching, and candidate feature correspondences are first computed based on the k-nearest neighbor algorithm. Then, all possible pairs of features are considered in terms of geometric deformation. In order to deal with pairwise geometric relationship between features, the spectral matching algorithm with pairwise constraints is applied. In the spectral matching process, a new robotic context-based correspondence filtering method is combined to obtain improved feature matching with less false correspondences. Finally, the RANSAC-based refinement is carried out to remove outliers. Also, we propose a new assessment method to obtain the final decision to accept or reject the matching results based on an affine distortion measure between model features and the matched image features. Experimental results show that the proposed object recognition algorithm can robustly identity less texture objects.