Stereo-based real-time scene segmentation for a home robot

This paper proposes a real-time scene segmentation method based on stereovision and intended for the use on a home service robot. In the first step of our approach the input disparity image is replaced by a lower resolution image. Its pixel disparity values are the result of building histograms over small neighbourhoods in the original image and selecting the maxima. This significantly reduces noise as well as wrong matches, and allows for meaningful local processing. Using constraints derived from the geometry and kinematics of the system camera - robot, ground plane pixels in the lower part of the image are selected. A least squares plane fit is applied to these points to determine the parameters of the ground plane and the rest of the pixels is labelled either as ground point or object point. Removal of the ground leaves single standing clusters of disparities corresponding to objects, which serves as both input for obstacle avoidance as well as object classification.