A Hybrid Kalman Filter Based Algorithm for Real-time Visual Obstacle Detection

摘要

In this paper we present a real-time scene reconstruction algorithm for mobile robots which is applicable for visual collision avoidance and online map building. Our method processes a sequence of images which are taken by a single camera mounted on a mobile robot. In contrast to similar monocular shape-from-motion algorithms, we combine two different approaches: A classic motion stereo approach and an algorithm for scene reconstruction based on extended Kalman filters. We show that the disadvantages of the classic stereo approach are compensated by the Kalman filter and vice versa. Our special method of initializing the Kalman filter leads to a faster convergence compared to other Kalman based approaches that use different methods for initialization. Moreover, we present a feature matching algortihm which is faster and more reliable than the widely-used KLT-Tracker in the domain of scene reconstruction.