Real-time multi-scale parallel compressive tracking

Abstract

Robust visual tracking is a challenging problem because the appearance of a target may rapidly change due to significant variations in the object’s motion and the surrounding illumination. In this paper, a novel robust visual tracking algorithm is proposed based on an existing compressive tracking method. The proposed algorithm adopts multiple naive Bayes classifiers, each trained under a different scale condition, to realize online parallel multi-scale classification. Further, each classifier was initialized by randomly generating different types of Haar-like features. By doing so, the robustness of the feature classification can be improved to obtain more accurate tracking results. To enhance the real-time performance of the visual tracking system, the formula of the naive Bayes classifier is studied and simplified to speed up the processing speed of parallel multi-scale feature classification. After acceleration via formula simplification and parallel implementation, the proposed visual tracking algorithm can reach a tracking performance of approximately 45 frames per second (fps) when dealing with images of 642 × 352 pixels on a popular Intel Core i5-3230M platform. The experimental results show that the proposed algorithm outperforms state-of-the-art visual tracking methods on challenging videos in terms of success rate, tracking accuracy, and visual comparison.