Sequential Monte Carlo-guided ensemble tracking

• Published in: PloS one Vol. 12; no. 4; p. e0173297
• Main Authors: Wang, Yuru, Liu, Qiaoyuan, Jiang, Longkui, Yin, Minghao, Wang, Shengsheng
• Format: Journal Article
• Language: English
• Published: United States Public Library of Science 11.04.2017; Public Library of Science (PLoS)
• Subjects: Accuracy; Animals; Artificial intelligence; Bayes Theorem; Bayesian analysis; Biology and Life Sciences; Branches; Classification; Color; Computer and Information Sciences; Computer science; Condensation; Data processing; Decomposition; Deer; Deformation; Engineering; Filters; Forests; Gating; Humans; Illumination; Image processing; Image Processing, Computer-Assisted; Information technology; Intelligence; International conferences; Internet; Kernels; Learning algorithms; Methods; Monte Carlo Method; Monte Carlo methods; Monte Carlo simulation; Motion; Neural networks; Neurocomputing; Objective function; Occlusion; Optical tracking; Pattern recognition; Physical Sciences; Research and Analysis Methods; Robots; Segmentation; Support Vector Machine; Support vector machines; Technology; Tracking; Tracking (Psychology); Vision; Visual perception; Visual stimuli
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0173297

A great deal of robustness is allowed when visual tracking is considered as a classification problem. This paper combines a finite number of weak classifiers in a SMC framework as a strong classifier. The time-varying ensemble parameters (confidence of weak classifiers) are regarded as sequential arriving states and their posterior distribution is estimated in a Bayesian manner. Therefore, both the adaptiveness and stability are kept for the ensemble classification in handling scene changes and target deformation. Moreover, to increase the tracking accuracy, weak classifiers including Support Vector Machine (SVM) and Large Margin Distribution Machine (LDM) are combined as a hybrid strong one, with adaptiveness to the sample scales. Comprehensive experiments are performed on benchmark videos with various tracking challenges, and the proposed method is demonstrated to be better than or comparable to the state-of-the-art trackers.