Model-based object tracking using stereo vision

Robust and accurate object recognition and tracking is a necessary pre-requisite for visual servoing tasks. This paper presents a new approach using disparity images provided by a stereo vision system from which 3D models are reconstructed. We use a segmentation procedure based on a simple region growing process to segment the non-dense 3D images, and compare it to a classical dense range image segmenter. The obtained regions are classified according to a simple model of the target object, and multiple candidates are handled over image sequences by application of the CONDENSATION algorithm. We then use a histogram based clusterization technique to identify the target object pose for visual servoing commands. Experimental results in localizing a door-handle show an accuracy of about 10 cm for the estimated position, while being robust to clutter and noise in the sensor images.