Robust elevator button detection in completely reflective surfaces

Abstract

Elevator button (EB) detection is a very important research issue in home service robotics area, this paper describes an effective method for EB detection with mirroring effects in completely reflective surfaces. Line features and other local features (such as corner, SIFT and color, etc) are extracted from stereo images to represent the EBs, since the EBs are almost in the complete reflective surface with mirror effect, so there are many ¿virtual¿ features, which are not in the plane of EB panel, in order to get only ¿real¿ features from the EB panel, homography matrix between two stereo images is applied to eliminate the ¿virtual¿ features effectively. After feature matching with the EB model, multiple interpretations with corresponding weights and uncertainties are generated as hypothesis of EBs position. Finally, the proposed particle filtering based interpretation fusion algorithm is applied to estimate the multiple EBs position robustly and sequentially. Experimental results in the real environment demonstrate that robust detection and convergence performance are achieved when confronted by serious mirror effects.