Using Active Vision to Simplify Perception for Robot Driving

Abstract

As mobile robots attempt more difficult tasks in more complex environments, they are faced with combinatorially harder perceptual problems. In fact, computation costs for perception can easily dominate the costs for planning in a mobile robot. Existing perception systems on mobile robots are potentially many orders of magnitude too slow for real-world domains. In this paper we show active vision at the system level can make perception more tractable. We describe how our planning system for a complex domain, tactical driving, makes specific perceptual requests to find objects of interest. The perception system then scans the scene using routines to search for these objects in limited areas. This selective vision is based on an understanding and analysis of the driving task. We illustrate the effectiveness of request-driven routines by comparing the computational cost of general scene analysis with that of selective vision in simulated driving situations.