Visual Tracking With Sparse Representation Based Discriminative Dictionary Learning

Visual tracking is an important subject in computer vision and it is crucial to applications such as video surveillance, motion recognition, traffic monitoring and robot vision. Due to the success of sparse representation theory, recent year has witnessed remarkable progress in visual tracking technics. Sparsity based tracking models show better performance in experiments, while there are still challenges caused by camera motion, appearance changes, lighting variations and occlusions. Generally there are three categories of tracking models, generative, discriminative and collaborative models. For generative model, tracking is formulated as searching for the most similar region to the target object within a neighborhood. For discriminative model, tracking is treated as a binary classification problem which aims at designing a classifier to distinguish the target object from the background. Collaborative method takes both advantages of these two models by combining them together.This paper proposes a modified generative tracking model which is improved with a structured discriminative dictionary to help distinguish the target from the background. Such structured dictionary is obtained through two steps. In the first step, the background templates are added in the dictionary and in the following step this dictionary is trained by FDDL algorithm to be structured and discriminative with respect to different classes. Using discriminative dictionary to represent the target samples, the proposed model can exploit not only the reconstruction error but also the representation coefficients in decision-making process, thus is more robust compared with the original generative model. Besides, unlike the general collaborative model, the proposed model doesn’t suffer from the computational cost of training an extra classifier, which may increase the model complexity.