Using fuzzy least squares support vector machine with metric learning for object tracking. (December 2018)

• Record Type: Journal Article
• Authors: Zhang, Shunli; Lu, Wei; Xing, Weiwei; Zhang, Li
• Is Part Of: Pattern recognition. Volume 84(2018:Dec.)
• Page Start: 112
• Page End: 125
• Publication Date: 2018-12
• Subjects: Object tracking -- Metric learning -- Fuzzy least squares support vector machine with metric learning(FLS-SVM-ML); Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2018.07.012
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A new FLS-SVM-ML algorithm is proposed. A two-stage iterative process is used to solve the FLS-SVM-ML problem. A tracking algorithm based on FLS-SVM-ML is presented. Experimental results demonstrate the state-of-the-art tracking performance. Abstract: Some researchers have introduced the fuzzy learning into tracking and the kernelized fuzzy least squares support vector machine (FLS-SVM) has achieved great success in building the appearance model. However, the kernel used in FLS-SVM is fixed, which may potentially limit the adaptivity to different conditions. In this paper, we introduce metric learning into the FLS-SVM classifier and propose a novel tracking method based on the combination of fuzzy learning and metric learning to address the above issue. First, we propose a new fuzzy least squares support vector machine with metric learning (FLS-SVM-ML) algorithm, which embeds metric learning into the FLS-SVM method and is used to learn the kernel in FLS-SVM adaptively. Moreover, we present a two-stage iterative optimization process to solve the optimization problem. Second, we apply the proposed FLS-SVM-ML method into tracking based on the FLS-SVM tracking framework. By introducing the metric learning, the FLS-SVM-ML method can be used to improve the adaptivity of the appearance model to different video sequences and different frames in the same sequence. Experimental results demonstrate that the proposed tracking method can achieve competitive tracking results and outperform many state-of-the-art methods in the benchmark datasets.