A context constraint and sparse learning based on correlation filter for high-confidence coarse-to-fine visual tracking

• Published in: Expert systems with applications 2025-04, Vol.268, p.126225, Article 126225
• Main Authors: Su, Yinqiang, Xu, Fang, Wang, Zhongshi, Sun, Mingchao, Zhao, Hui
• Format: Article
• Language: English
• Subjects: Context constraints; Discriminative Correlation Filters; Lasso regression; Sparse learning; Visual tracking
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2024.126225

Discriminative Correlation Filters (DCFs) have recently garnered significant considerable in the field of visual single tracking. However, existing trackers frequently struggle to fully mine the structural complementarity and diversity among various features, resulting in a decline in discriminability in complex scenarios. To address these challenges, we explore a novel context-aware sparse learning (CCSL) framework based on DCFs and hierarchically infer the maximum response represented by each class filter to locate the target confidently. Guided by sparse learning principles, our tracker collaboratively and bidirectionally selects effective spatial elements that encode target appearance through lasso regression and sparse response estimate. Additionally, the target and its surroundings are jointly incorporated into the learning framework, thereby bolstering the discriminability of the learned model. To tackle the optimization problem, we leverage the Alternating Direction Method of Multipliers (ADMM) in the Fourier domain. Furthermore, we introduce a hierarchical inference scheme for target localization that harnesses complementary cues from different features, which leverages historical displacement and multi-modal detection to reveal the tracking state respectively. Extensive experiments are conducted on renowned benchmarks, including OTB-100, UAV123, UAV20L, and TC128, to demonstrate the efficiency and effectiveness of the proposed tracker. These results underscore the tracker’s ability to perform robustly in complex scenarios, showcasing its superior discriminative power and accuracy.