Research Of Object Tracking Based On Local Appearance Model

Object tracking plays an important role in the field of computer vision, with the goal of lo-cating the target and capturing the change of target’s appearance and motion in each frame. This paper summarizes classic theories and algorithms in the field of object tracking, and gives com-prehensive analysis on different challenges in object tracking task, state-of-the-art approaches and the history and development trend of this area. Two kinds of object tracking methods based on local appearance model are proposed in this paper, that is, Visual tracking via adaptive struc-tural local sparse appearance model and Fragment-based tracking using online multiple kernel learning.This paper first introduces the local appearance model into the generative tracking frame-work and proposes a structural local sparse appearance model, which is able to exploit both local and spatial information of the target via sparse coding on local patches, and averaging and pool-ing operations on the obtained coding coefficients. This helps the tracker make full use of stable local patches within the target region to carry out robust object tracking. In addition, an update strategy, which is based on sparse representation and incremental subspace learning, is proposed to adapt the appearance change of the target with less possibility of drifting and reduces the influence of the occluded target template as well.The local appearance model is then introduced into the discriminative tracking framework and a novel fragment-based tracking approach using multiple kernel learning method (MKL) is proposed. An online MKL method for object tracking is developed by considering temporal con-tinuity in video explicitly. This method can fuse discriminative ability of fragments by assigning them different weights according to their discriminative power. In addition, for better robustness two kinds of independent features are computed to enrich the representation of the object. One classifier is trained for each type of features and they are assigned different weights according to their performance.Both qualitative and quantitative evaluations on challenging image sequences demonstrate that the proposed tracking algorithms perform favorably against state-of-the-art methods.