Robot object recognition based on learning and cognition with conceptual space

In view of the challenging issue for robot multi-class object recognition in complex environments with original feature space because those features can only embody some low level knowledge with poor discriminative performance, an approach to robot object recognition was presented based on conceptual space methodology. At first, the basic feature space was built through multi-sensor data fusion and feature extraction, thus gaussian mixture models (GMM) was employed to model objects properties in order to build conceptual space with characteristic of high level knowledge so as to learn the concepts of objects and improve their discriminative performance. Then support vector machine (SVM) was used to carry out multi-class object recognition for robot in indoor clutter environment. The experimental results demonstrate that the proposed method can not only represent the high level knowledge of objects, but also improve the performance of object recognition and environment perception of robot effectively.