Research On Visual Tracking Algorithms Based On Discriminative Model

As the code of the high level research paper is released as open source, and the benchmark datasets is increasingly expanded, visual tracking research has made rapid development in recent years. Although these work are very valuable, whether they are sufficient for understanding and diagnosing the advantages and disadvantages of the different trackers remains questionable. In order to address this issue, we propose a framework that can diagnose the visual tracking algorithms, by breaking a tracker down into five parts, namely, training samples acquisition module, the observation model module, testing samples acquisition module, feature representation module, and observation model updating module. We then conduct ablative experiments on each component to study how, and to which extent it affects the overall performance of tracking system separately.Particularly, in view of each module, this paper puts forward the corresponding improvement and setting. In the training samples acquisition module, we use the nonnegative least squares,Bhattacharyya coefficient and Bhattacharyya distance algorithms to weight labeles of samples;In the observation model module, we mainly focus on Ridge Regression,since it admits a simple closed-form solution, but a low computing complexity, and can achieve performance that is close to more sophisticated methods, such as Support Vector Machine; In testing samples acquisition module, we use particle propagation model to collect testing samples in particle filter algorithm;In feature representation module, we propose a novel color feature, and connect it with the histogram orient of gradient feature directly to develop a stronger combination feature representation;In the observation model updating module, we only summarizes the existing two main methods.Based on our improvements, this paper put the improved modules together to generate a new tracker again and compare its performance to the top 10 trackers in the benchmark.Finally we find that this new constructed tracker is competitive in performance to the state-of-the-art trackers.