This article describes a method for vision-based autonomous convoy dri
ving in which a robotic vehicle autonomously pursues another vehicle.
Pursuit is achieved by visually tracking a target mounted on the back
of the pursued vehicle. Visual tracking must be robust, since a failur
e leads to catastrophic results. To make our system as reliable as pos
sible, uncertainty is accounted for in each measurement and propagated
through all computations. We use a best linear unbiased estimate (BLU
E) of the target's position in each separate image, and a polynomial l
east-mean-square fit (LMSF) to estimate the target's motion. Robust an
tonomous convoy driving has been demonstrated in the presence of vario
us lighting conditions, shadowing, other vehicles, turns at intersecti
ons, curves, and hills. A continuous, autonomous, convoy drive of over
33 km (20 miles) was successful, at speeds averaging between 50 and 7
5 km/h (30-45 miles/h).