Robust visual tracking based on local kernelized representation
Qiaozhe Li, Yu Qiao, Jie Yang · 2014
Visual tracking, when regarded as a classification problem, usually need sufficient amount of data for online learning. Current approaches usually employ multiple binary labeled samples to train or update the appearance model at each frame. However, most positive samples collected by perturbation of the target location are equally treated as the estimated target when used for updating. The exact target information might be ambiguous. In this paper, we propose a robust visual tracking algorithm based on local kernelized representation. The use of kernels incorporates the target information to better represent each samples. Local kernel scores, which measure the similarity between samples and target templates in specific feature space, are used to form feature vector to represent the entire region. A neural network is then deployed into the particle Alter framework to estimate target location. Experimental results on a variety of challenging sequences demonstrate that the proposed tracking algorithm performs favorably against several state-of-the-art methods.