Adaptive Processing for Vision-Based Ranging

Abstract

Consider a camera that moves through a stationary environment. The time-varying range to features in the camera's field of view may be estimated recursively using a sequence of images. We study processing these images adaptively; our goal is accurate range estimation and increased range map density in the presence of constrained computational resources. The problem of autonomous nap-of-the-earth rotorcraft flight motivates our work. An onboard camera and inertial measurement unit provide sensing for the construction of local range maps for an obstacle avoidance function. To provide more features yielding accurate range estimates, we assume a high resolution, high data rate camera. The resulting explosion in the image measurement data rate can be handled by selecting which measurements to process, and by making the best use of onboard computation in processing these measurements. Computer vision research provides many ideas on selecting which measurements to process. We study how to process the measurements. Specifically, for a single feature, we explore the idea of adaptively trading spatial resolution and image rate to obtain accurate range estimation while simultaneously satisfying an upper limit on computation rate. Simulations are presented to explore the effectiveness of this idea. Our work is applicable to other vision-based ranging problems.