| The goal of this thesis is to investigate and implement novel computer vision techniques based on active vision, that dynamically manipulate cameras (mounted on autonomous robots) in order to better explore and understand the environment, as compared to static camera solutions. Such an active vision approach can help with: (i) avoiding occlusion by direct camera manipulation, (ii) recognizing objects of interest from different viewpoints and with different levels of camera resolution, (iii) dynamically selecting the best source of sensor data for a specific recognition task, and (iv) seamless integration of data from multiple sensors.;In designing a perception mechanism for a robotic system, employing multiple sensing capabilities is beneficial for increasing the recognition abilities of the robot. In this thesis we design an active vision system with dynamic camera allocation, and we implement it on a PR2 robotic platform. It involves a head-mounted 3D camera (capturing RGB and depth information) working in cooperation with a hand-mounted camera (capturing RGB information only). The hand camera is dynamically employed to assist in recognizing objects for which the head camera cannot provide a confident classification, by matching the detected objects and fusing the recognition results. Our experimental results show considerable improvement over single camera object recognition. |
