Options
2005
Conference Paper
Titel
Stereo image based collision prevention using the census transform and the snow classifier
Abstract
In this paper we show an approach for a stereo-camera based system on a moving roboter to avoid obstacles. We propose the "census transformation" for generating the feature for the correspondence search. We train two SNOW (spare network of winnovs)-classifiers, one for the decision whether to move straight forward or to evade and a second one for deciding whether to turn left or right when evading. For training we use a sample set collected by manually moving around with the robot platform. We evaluate the performance of the whole recognition chain (feature generation and classification)using ROC-curves. Real world experiments show the mobile robot to safely avoid obstacles. Problems still arise when approaching steps or low obstacles due to limitations in the camera setup. We propose to solve this problem using a stereo camera system capable of pann and tilt movements.