Abstract
Recently, the service robots that work in human living environment are developed. At such environment, there are a lot of environmental changes. In addition, various color and shape objects exist at cluttered environment. So, it's important to find and recognize various target objects at partially occluded scene. To achieve that, we adopted part-based model. However, part-based model often causes false positive. In order to prevent that, we propose two-stage learning method. First stage is just part-based model. Second stage is the model trained from the tendency for part-based model to recognize wrong. By our method, the robots enable to handle partial occlusion. In addition, we use RGB-D features to improve the recognition accuracy of part-based model. Finally we achieve high performance at partially occluded scene for large dataset. And, our method can perform at 1.7[fps].
