Kernel-Based Tracking Using Online SVM Model Updating And Particle Swarm Optimization.

Abstract: The traditional MS tracker has two main drawbacks: (1) The template model can only be built from a single image; (2) It is difficult to adaptively update the template over the tracking. In this paper we attempt to use an online SVM based kernels model updating method to overcome these above-mentioned problems. Simultaneously, in contrast with the template matching method which one usually achieves tracking through maximizing the likelihood between model and candidate region. Here, we treat the tracking as a foreground/background classification problem. Initialized with a small number of data to train using SVM, we could get primitive models (include the object models and the background models) and a decision function. In the process of tracking, we update the previous models using the tracked data, and then maximize the value of decision function with the help of the particle swarm optimization. The optimal solution predicts the object''s location. Experiments show that this method performs better than MS tracker on some challenging sequences. [Copyright &y& Elsevier]