Autonomous three-dimensional tracking for reconfigurable active-vision-based object recognition

Abstract

Recognition algorithms would significantly benefit from object images acquired from preferential view points, e.g. unobstructed frontal views or complementary views. Active-vision systems, which are dynamically reconfigurable in an online mode, have been suggested in the literature as effective solutions for achieving this objective, namely, relocating cameras to obtain optimal visibility for a given situation. To obtain optimal visibility of an object of interest (OI), however, that OI's three-dimensional (3D) position and orientation (i.e. six degree-of-freedom pose) must be tracked in real time. Thus, this paper presents such an autonomous, real-time, six degree-of-freedom pose tracking system for a priori unknown objects. The proposed tracking method autonomously (a) selects the OI, (b) builds its approximate 3D model and uses this model to (c) track it in real time. As will be shown in this paper, via experimental results, the output of the proposed modeller can be effectively used by an active-vision system to relocate its cameras for effective preferential image acquisition. In the examples included herein, it will be noted that object visibility data obtained via camera reconfiguration based on the authors' ‘approximate’ tracking method are comparable with those obtainable based on ‘perfect’ OI tracking.