Siamese Deep Q-Learning Based Online Correlation Filter Adaptation for Visual Object Tracking in Complex Scenarios

Visual Object Tracking (VOT) is essential for real-time monitoring and analysis across various applications, but Siamese tracker networks face significant challenges, including overfitting, occlusion handling, fast motion, complex backgrounds, and adaptation to appearance changes. To address these challenges, we propose a novel tracker model named Siamese Deep Q-learning based online Correlation Filter Adaptation (SiamDQCFA). SiamDQCFA uses two identical subnetworks to capture evolving appearance changes and semantic similarities between the target object and search region patches, enabling effective feature extraction and adaptive parameter adjustment. This architecture extracts high-level feature representations that predict the target’s location in subsequent frames, improving accuracy over time. The model evaluates target presence in the search region using cosine similarity as a measure. Furthermore, SiamDQCFA incorporates an online correlation filter adaptation mechanism that dynamically adjusts its parameters to enhance target localization, improving the robustness of the tracker to various environmental challenges. Through extensive experimental evaluations on benchmark datasets such as OTB100, TrackingNet, VOT2019, and LaSOT, SiamDQCFA demonstrates superior performance, outperforming state-of-the-art tracking methods across multiple metrics. These results underscore the model’s effectiveness in addressing the key challenges of VOT, which makes it a robust solution for real-time tracking in challenging conditions.