ACTIVE STEREO VISION FOR MANIPULATOR ROBOT

Abstract

This thesis describes a novel active stereo vision platform that was developed to provide visual guidance to fruit harvesting robots. A five degree of freedom camera platform was built in order to identify and localize tomatoes grown in greenhouses in farms. A novel cognitive model of attention with three main features map is used to control the gaze of the camera system. The three feature maps are (A) an online epipolar geometry update to produce a disparity map and affordance of grasping, (B) a vergence controller to provide a map of the depth of tomatoes at high accuracy and (C) a bottom-up saliency map to detect and identify ripe tomatoes. The properties and performance of each of these maps is studied in detail. The final system, tuned to the visual properties of the tomato, has been assessed and reveals a detection performance of 82% and operates over a range of 200+/-0.2cm. The proposed system is not limited to tomato detection, since the attention system can be adjusted to different type of crops such as apply, orange etc. The camera system is designed to integrate with any robotic manipulator arm.