A robust landmark-based system for vehicle location using low-bandwidth vision

摘要

This paper presents novel computer algorithms, a system architecture, and the prototypeimplementation of a vision-based automatic vehicle location system. The objective of the vehiclelocation system is to keep track of the vehicle location for a human driver, and perhaps to provide thedriver with real-time audio directions to his destination. The techniques developed here are equallyapplicable to autonomous robot navigation. The prototype system uses odometer readings and askeleton map to perform dead reckoning, and uses low-bandwidth visual information and neuralnetworks to recognize places for correcting cumulative dead reckoning errors. The visual informationis also used to detect turns, for dead reckoning at intersections. The system is self-contained in thesense that it requires no infrastructure outside the vehicle, such as external beacons installed onroadways or satellites used by Global Positioning Systems (GPS). The system maintains a largenumber of location hypotheses and searches for a large number of landmarks stored in a database inreal time. Hence the system is robustly able to recover the vehicle location after being lost for variousreasons. The system has been tested, with success, in both day and night time, in all four seasons, andon roads in New York City, a regional highway, and on suburban streets.