Stereo Vision Based SLAM Issues and Solutions

Abstract

Simultaneous Localization and Mapping (SLAM) has been one of the active research areas in robotic research community for the past few years. When a robot is placed in an unknown environment a SLAM solution attempts to build a perfect map of the environment while localising the robot with respect to this map simultaneously. Traditionally SLAM utilised endogenous sensor data in the process. Successful SLAM implementations using laser (Guivant and Nebot, 2002), sonar and radar (Clark and Dissanayake, 1999) can be found in the literature, which prove the possibility of using SLAM for extended periods of time in indoor and outdoor environments with well bounded results. Recent extensions to the general SLAM problem has looked in to the possibility of using 3-dimensional features and the use of alternative sensors to traditionally used lasers and radars. Cameras are competitive alternatives owing to the low cost and rich information content they provide. Despite the recent developments in camera sensors and computing, there are still formidable challenges to be resolved before successful vision based SLAM implementations are realised in realistic scenarios. Monocular camera based SLAM is widely researched (Davison et al., 2004; Kwok et al., 2005), however, binocular camera based SLAM is mostly overlooked. Some of the noted stereo implementations can be found in (Davison and Murray, 2002) and recently in (Jung, 2004). Lack of enthusiasm for research in this direction could possibly be attributed to the misconception that range and bearing information provided by the stereo vision system is directly utilizable providing a simplistic solution to SLAM which is academically less appealing or the apparent success in single camera SLAM implementations. However, after rigorous analysis and sensor modelling, we found that the standard extended Kalman filter (EKF) based SLAM with small base line stereo vision systems can easily become inconsistent (Herath et al., 2006a) . This chapter attempts to provide readers with an understanding of the SLAM problem and its solutions in the context of stereo vision. The chapter introduces the Extended Kalman Filter as applied to the generic SLAM Problem. Then, while identifying the prevailing issues inherent in solutions to the SLAM problem in stereo vision context, our solutions are presented with simulated and experimental evaluations. Several components of the stereo