Tracking in cluttered images

摘要

A new algorithm, the competitive attentional tracker is proposed, which combines multiple velocity estimating filters to detect and track targets in cluttered images. Each filter tracks the motion of some scene content using a discrete grid representation for position and velocity beliefs. During operation, the filters can move between high confidence states (accurate tracking) and low confidence states (maximum detection sensitivity). Optimal detection performance is related to a hidden Markov model for target presence/absence, which can exploit statistical models of target and background appearance to detect targets in highly cluttered scenes. Experiments with synthetic data are used to characterise detection and tracking performance, and implementation in target tracking systems is described.