Dynamic confident threshold strategy for low-slow-small UAV detection task based on KDE

In order to solve the problem of frequent target loss and redetection when using deep learning to carry out visual detection of low-slow-small UAVs which travel between complex backgrounds, this paper constructs a set of local confidence threshold adjustment mechanism and uses particle swarm optimization (PSO)-random forest(RF) and kernel density estimation(KDE) algorithms to make interval prediction of UAV trajectory and obtain an optimal estimation of its future distribution region, and subsequently lowering the confidence threshold within this region. Compared with fixed confidence detection methods, the proposed adjustment mechanism effectively reduces missed detection rates of UAV video targets under complex backgrounds.