Robust Tracking Via Multi-level Multi-feature Templates And Adaptive Dynamic Model

Object tracking is an essential topic in computer vision and has significant research value. Tracking algorithm has been implemented in various visual applications, such as video surveillance, video compression, and human-computer interface. Therefore, it is required to deal with multiple practical challenges like illumination change, object scale variant and rotations. Finding the way to improve tracking accuracy and robustness still remains a difficult challenge in tracking area.This paper focuses on object appearance model and motion estimation subarea in object tracking. We proposed the tracking algorithm based on Multi-level Multi-feature Templates Model(MMTM) and tracking algorithm based on Adaptive Dynamic Mode(ADM), respectively.Firstly, to improve the object appearance model, a robust tracking algorithm via MMTM is proposed. In this algorithm, an object is represented by a hierarchical tree model where the nodes represent different object parts. Specifically, a three-level spatial pyramid is utilized to characterize the target in high-level, mid-level and low-level representations. An appearance template is designed for each part and is described by a multi-feature set that combines multiple visual observations, such as color, intensity, texture and edge. We demonstrate the MMTM for object tracking in a generative learning formulation based on information projection. In this formulation, the weights of different features in each template and the contribution of different templates are learnt by maximizing the difference between underlying probability model for the tracked object and the probability distribution of background images. The proposed model is tested on the currently largest benchmark, and achieves higher tracking accuracy(0.7 ~ 7.2 percent) than CVPR 2012 proposed methods. Experimental results reveal that this method is extremely robust to challenges like background cluttering, object rotation and motion blur.Secondly, we dug into motion estimation area and proposed a tracking algorithm based on ADM. In order to improve motion estimation accuracy and robustness, this algorithm introduced adaptive model degeneration to the first-order particle filter, and the degenerated model is the random walk model. The ADM algorithm decides the degeneration degree based on the predication accuracy of the first-order dynamic model, and meanwhile changes the parameters for both the first-order model and the random walk model adaptively. Degeneration degree is represented by the particle number distribution in first-order and random walk model. The adaptively changed parameters in these two models are the variances of their evolution noise. Experimental results demonstrate the ADM is more accurate and more robust than the traditional first-order dynamic model in object tracking. The experiment also reveals that, compared to the original first order particle filter model, applying ADM to the tacking algorithm will greatly improve tracking accuracy(6.6 percent). And compared to the pyramid searching method, ADM has better detecting efficiency.