Hybrid Particle Filter and Mean Shift tracker with adaptive transition model

Appearance learning by adaptive Kalman filters for FLIR tracking

“…A commonly used distance measure is the Bhattacharyya Coefficient (BC) [2,7]. However, we have found that the histogram intersection metric [16] is more suitable for the problem considered here.…”

“…The lack of color information in IR images prohibits the use of color features as in [2]. Because they are scale invariant and tend to be slowly varying, intensity histograms are widely used for target representation [2,7,21]. Here, we employ a dual foregroundbackground appearance model [17] which incorporates simple pixel statistics (intensity and local stdev) from both the target and the surrounding background.…”

Accurate appearance-based Bayesian tracking for maneuvering targets

“…A different approach to particle sampling is to drive the particles according to point estimates of the gradient of either the posterior or the likelihood [21,20,24,25,28,29]. When the appearance model is a template, optical flow can be used to drive the particles towards peaks of the likelihood.…”

“…For this reason, color-based MS fails to track small and fast moving targets as well as to recover the position of a target after a total occlusion. Given the complementarity of the two algorithms, combinations of MS and PF have been studied [28,29]. However, the convergence of the MS procedure with the used flat kernel is not demonstrated [28].…”

“…However, the convergence of the MS procedure with the used flat kernel is not demonstrated [28]. Moreover, as the particles are re-displaced by MS, it is not clear how the new sampling distribution can be approximated to correctly compute their weights [28,29].…”