Object Recognition based on Depth Aspect Image Matching
Tomoyuki Takeguchi, Tsukasa Kondo, Shun’ichi Kaneko, Satoru Igarashi · 2000
A method for three dimensional object recognition based on depth image information is proposed. A depth aspect image is defined as an orientation standardized appearance from the original depth data of the object, which is transformed by the rigid transformation drawn by each possible basis pair of every three feature points of the object depth data. They are made from the original depth images of models and then learned in the system as the database for retrieval of any instances on the models. Matching between an object aspect and the ones from models can be performed by two-dimensional image comparison, which is based on the least quantile of residuals and is robust against occlusion possibly occurred in cluttered scene. The paper includes a formalization of the proposed method and some experimental results with real objects. fine resolution of partial shapes of models[8] A novel matching scheme is also presented, which is based on a robust statistic and hence model instances with occluded part can be searched in the real scene. Depth aspect images are fundamental in the method and they are made through relative coordinates by the feature points on the model surfaces, enabling position and posture estimation of the models in the scene. The proposed method is suitable for hardware realization and so fast and real time processing. The paper consists as follows: In Section 2, an outline of the method is given together with definition of depth aspect images and the algorithm for generating them. In Section 3, a robust recognition algorithm using image-based matching is given for handling complex scenes with multiple objects. In Section 4, experiments with real scenes are pre sented, and then we conclude the paper with some remarks in Section 5. 1 Introduction 2 Depth