Adaptive Perception for Autonomous Vehicles

Abstract

A new approach to autonomous vehicle perception is presented which solves the historically significant throughput problem at contemporary speeds through computational stabilization of the sensor sweep. This adaptive approach to perception has made it possible to achieve unprecedented autonomous vehicle speeds at little or no cost to other aspects of performance. In order to measure the local environment at sufficient resolution and sufficient rate, an autonomous vehicle requires computational throughput on the order of O [TV2] where T is the vehicle reaction time and V is the velocity. On the other hand, the traditional approach of nonadaptive range image processing requires throughput on the order of O [T4V4]. The product TV is on the order of 10 for a conventional automobile so the difference between these two expressions is four orders of magnitude at 20 mph. Nonadaptive range image processing requires about 1 gigaflop in order to achieve 20 mph speeds whereas the algorithm presented here requires 1/10 of 1 megaflop under identical assumptions. This report concentrates on the adaptive perception algorithm which forms the basis of RANGER's Map Manager object. The techniques used should be applicable to any application that models the environment with a terrain map.