Computer vision for CARMEL

In this paper we discuss the implementation and uses of the object recognition system used for CARMEL, the University of Michigan's winning entry in the AAAI{92 Autonomous Robot Competition. Following the rules of the competition, the robot was required to navigate within a large, unstructured environment performing exploration, and then a directed search, for objects placed throughout the arena. CARMEL was completely autonomous and performed these tasks, in part, using computer vision techniques. The tasks required of the computer vision system consisted of actively searching for objects (four inch diameter tubes marked with black and white stripe patterns), detecting them in images, uniquely identifying each object based upon its distinguishing pattern, and determining each object's position from orientation and distance estimates measured from the image. We brie y describe the design of the various computer vision algorithms that were developed to perform these tasks. Because of the accuracy and robustness of the vision system, we were able to perform absolute positioning, where the robot accurately updated its position through backward triangulation from previously located objects. The success of CARMEL stemmed largely from the use and implementation of the vision system to perform the tasks listed above. Other teams chose to approach these same tasks using dierent sensory systems and/or techniques. We analyze the general approaches, looking at where they excelled and failed, in terms of their actual performance and in general, perhaps giving insight into how to build autonomous robots that can successfully operate in \natural " environments