Comparison between Kalman and unscented Kalman filters in tracking applications of computational vision

Abstract

In this paper, the problem of tracking feature points along image sequences is addressed. The establishment of correspondences between points and their tracking along image sequences is a complex problem in Computational Vision; especially, when intricate motions, erroneously detections or cases of occlusion or appearance/disappearing of features are involved. To overcome some of those difficulties, a statistical ap-proach is frequently used in a multi-object data association and state estimation framework. Additionally, the correspondence between each measurement and predicted feature can be performed by minimizing the overall Mahalanobis distance. Under these circumstances, the estimation of the system can be accomplished using different stochastic filters. Hereby, a comparison is made between the results obtained, with the described framework, either by the Kalman Filter or the Unscented Kalman Filter, in the tracking of linear and non-linear motions of feature points along image sequences.