We use multiple tracking agents in parallel for autonomously tracking an arbitrary target from an unmanned aerial vehicle.
An object initially selected by a user
from a possibly cluttered scene containing other static and moving objects and
occlusions - both partial and complete - is tracked as long as it remains in view
using a single light-weight camera readily installed in a UAV.
We assumed, for the present, at least, that the UAV sends images
to a ground station which controls it.
We evaluated several individual tracking agents in terms of tracking success and their times for processing
frames streamed from the UAV to the ground station at 25 fps, so that the system
shoud compute results in 40ms. Histogram trackers were most successful at $\sim 10$ ms per frame
which can be further optimized.
Ultra low-power visual odometry for nano-scale unmanned aerial vehicles
