CONTROL OF PERCEPTUAL ATTENTION IN ROBOT DRIVING

Computer vision research aimed at performing general scene understanding has proven to be conceptually difficult and computationally complex. Active vision is a promising approach to solving this problem. Active vision systems use optimized sensor settings, reduced fields of view, and relatively simple algorithms to efficiently extract specific information from a scene. This approach is only appropriate in the context of a task that motivates the selection of the information to extract. While there has been a fair amount of research that describes the extraction processes, there has been little work that investigates how active vision could be used for a realistic task in a dynamic domain. We are studying such a task: driving an autonomous vehicle in traffic. In this paper we present a method for controlling visual attention as part of the reasoning process for driving, and analyze the efficiency gained in doing so. We first describe a model of driving and the driving environment, and estimate the complexity of performing the required sensing with a general driving-scene understanding system. We then introduce three programs that use increasingly sophisticated perceptual control techniques to select perceptual actions. The first program, called Ulysses-1, uses perceptual routines, which use known reference objects to guide the search for new objects. The second program, Ulysses-2, creates an inference tree to infer the effect of uncertain input data on action choices, and searches this tree to decide which data to sense. Finally, Ulysses-3 uses domain knowledge to reason about how dynamic objects will move or change over time; objects that do not move enough to affect the robot's decisions are not selected as perceptual targets. For each technique we have run experiments in simulation to measure the cost savings realized by using selective perception. We estimate that the techniques included in Ulysses-3 reduce the computational cost of perception by 9 to 12 orders of magnitude when compared to a general perception system.