An interactive open-ended learning approach for 3D object recognition

Abstract

Three-dimensional object detection and recognition is increasingly in manipulation and navigation applications in autonomous service robots. It involves clustering points of the point cloud from an unstructured scene into objects candidates and estimating features to recognize the objects under different circumstances such as occlusions and clutter. This paper presents an efficient approach capable of learning and recognizing object categories in an interactive and open-ended manner. In this paper, 'open-ended' implies that the set of object categories to be learned is not known in advance. The training instances are extracted from actual experiences of a robot, and thus become gradually available, rather than being available at the beginning of the learning process. This paper focuses on two state-of-the-art questions: (1) How to automatically detect, conceptualize and recognize objects in 3D unstructured scenes in an open-ended manner? (2) How to acquire and utilize high-level knowledge obtained from the user (e.g. category label) to improve the system performance? This approach starts with a pre-processing phase to remove unnecessary information and prepare a suitable point cloud. Clustering is then applied to detect object candidates. Subsequently, all object candidates are described based on a 3D shape descriptor called spin-image. Finally, a nearest-neighbor classification rule is used to assign category labels to the detected objects. To examine the performance of the proposed approach, a leave-one-out cross validation algorithm is utilized to compute precision and recall. The experimental results show the fulfilling performance of this approach on different types of objects.