Stereoscopic Vision for Off-Road Intelligent Vehicles

Background

Although the principles of stereoscopy have been known since the nineteen century, the availability of commercial stereo cameras only dates from the turn of this century. The processing speed of current computers allows the execution of algorithms that can correlate two stereo images and generate a depth map in real time. The level of detail and amount of information supplied by stereoscopic perception has placed stereo-based devices in a privileged position among other sensors used in field robotics. Mars exploration (Olson et al., 2003) and defense mobile robots like Urbie rely on stereo cameras to acquire critical information of remote and often hazardous environments. The application of 3D vision technology to agricultural vehicles, in spite of having a high potential (Rovira-Más, 2003), is still in its infancy. Some timid efforts have been made to apply the idea of stereoscopy to automatically locate fruits in plants (Kondo et al., 1996), but human intervention has been normally required to assist in pixel matching. Real time stereo-based perception for mobile robots is relatively recent, and although some solutions have been successfully developed for small indoor robots (Herath et al., 2006) and on-highway vehicles (Kato et al., 1996), the scenarios typically perceived in these applications are substantially different from those encountered by off-road vehicles; therefore, the latter demand specific solutions motivated by very distinctive needs. Even the off-road prototypes that participated in the DARPA Grand Challenge competition, organized by the United States Department of Defense, were set to fulfill elaborated missions that nothing have to do with habitual agronomical tasks (Kogler et al., 2006).