Stereo vision and mapping with unsynchronized cameras

Abstract

Environmental awareness is an important prerequisite for autonomous behavior in vehicles. Without it, robots are unable to react to unknown surroundings and require extensive human input for tasks such as target identification and obstacle avoidance. This would negate many of the advantages of having an autonomous system. Giving a vehicle the ability to map its surroundings and use the data effectively will allow humans to spend less time scanning the vehicle's video feed and providing direct navigational commands. This thesis details the development of a real-time, extensible vision and mapping system that provides an interface for control systems to access details of the map. It addresses the problems of image capture, signal noise, and three dimensional map storage. It extends existing real-time stereo mapping systems by tolerating unsynchronized stereo cameras. Results indicate that synchronization allows the system to locate points significantly more accurately than the system without synchronization. When compared with a monocular mapping system, synchronized stereo provides a more detailed map and will tolerate more erroneous localization data. Because it is developed with an abstract localization system, this system is designed to be modular and easily extensible.