Real-time pose estimation and control for convoying applications

Abstract

This results support our position that by putting together traditional computer vision techniques carefully customized to the to meet application-specific needs, it is possible to tackle challenging problems with low-cost off-the-shelf hardware. In the specific case of convoying, we have shown, in a careful evaluation with synthetic data, that specialized motion analysis algorithms that take into account domain-specific constraints such as the existence of a unique ground plane often yield more accurate and stable results than totally generic techniques, even when these assumptions are only partially met. Finally, we suggested a two-level approach for control, in which high-frequency odometry data is used to stabilize visual control.

This paper describes work that is still in progress and we stress the fact that some of the issues raised here need further investigation. In our opinion one of the most interesting directions in which this work must be continued is with a deeper investigation of which is the best control strategy for the application at hand. Our current controller assumes “off-road” conditions: it is permissible always to head directly at the lead vehicle, thus not necessarily following its path. If vehicles must stay “on road”, the follower may be forced to re-trace the trajectory of the leader precisely. State estimation of the leader’s heading (global steering angle, say) as well as speed (or accelerations) are ultimately needed, to be duplicated for local control. Vision becomes harder since the follower cannot always aim itself at the leader. The desired trajectory is known, which turns the problem into one that can perhaps more usefully be related to optimal control than to simple feedback control.