The Kalman-EM Contour Tracker

Many active-contour trackers include separate modules for feature detection, pose/shape refinement, and Kalman filtering. While this design is convenient for the programmer, it is not statistically optimal, because (a) feature detection leads to information loss and (b) the pose/shape refinement module makes no use of prior information from the filtering module. The tracker introduced in this paper employs the Bayesian EM algorithm for pose refinement: applying the EM algorithm to active contours eliminates the need for a separate feature-detection stage, while the Bayesian form of the EM algorithm provides a natural way to incorporate a prior estimate (from the filtering stage) into the pose refinement stage. Performance comparisons show that the Bayesian EM Contour algorithm is faster and more robust than modular contour trackers. The advantage of the Bayesian EM Contour algorithm over particle filtering is demonstrated by the much smaller number of function evaluations needed to achieve comparable tracking performance.