Object Tracking Based On Patch Modeling And Online Learning

In many areas, there are lots of important applications of detectingand tracking for moving targets in the image sequences， such asintelligent transportation systems, medical diagnostics, remote sensingcontrol and man-machine interaction, and so on. The tracking ofnon-rigid targets is full of the uncertainty and complexity in the shape,the appearance and the structure of the target. How to improve therobustness and adaptability of the algorithms for tracking non-rigidtargets is an important issue to the researchers. While how to keepsatisfying tracking results of temporarily disappearing objects is anothersignificant issue.This paper studies TLD (tracking-learning-detection) and thetracking algorithm of a non-rigid object via patch-based dynamicappearance modeling and adaptive basin hopping Monte Carlo sampling.TLD combines the tracking and detecting by P-N learning, which hasgood tracking results for the target of which geometric appearancechanges gradually or it is of a brief disappearance, but it is not suitablefor tracking non-rigid objects. The improved particle filter based on patchmodeling is applicable for non-rigid objects tracking. While it fails tore-track the object when it is of a brief disappearance, as the particle filterhas no ability of online learning.This paper improves the two algorithms, proposing a new one basedon the patch modeling while tracking the patches by the online learningof TLD. This paper is supposed to put forward the initialization methodof the patches based on the condition number of the Hessian Matrix, the updating method of patch degradation and the target reduction strategy,valuing and modifying the patches for finally restoring the whole object.The experiments show that this algorithm can well track the rigid objectswith different velocity and has higher accuracy than TLD for non-rigidobjects of which geometric appearance change drastically over time.Even if the target is of a brief disappearance, it can still be recovered bythis algorithm, while the patch-based particle filter fails to.