Abstract
The purpose of this study is to estimate the position and orientation of a chair that is put into a desk. In order to estimate the position and orientation of the chair, we use a depth camera which provides point clouds of the object. First, we extract three-dimensional local feature descriptors called "FPFH" from the point clouds. Second, we calculate Bag-of-Features from FPFH descriptors and label each point in the cloud. After that, by considering the arrangement of the labels, we try to find the chair from the point clouds of several objects include the desk and the chair. We have carried out experiments, targeting five chairs that are put into the desk. As a result, we found that our approach is effective to recognize the position and orientation of four types of chairs among the five, in approximately 60-90 % success rate. However, because of occlusion arising from a complicated shape of the remaining one type chair, the recognition was not successful only with a 25 % success rate.
