Scholarlink

<p>Although visual object tracking (VOT) has achieved significant milestones, robust occlusion handling remains a formidable challenge, often acting as the primary catalyst for identity switches and tracking failures. While existing related surveys predominantly focus on broad tracking architectures or general deep representation learning, they often treat occlusion as a secondary component and function as annotated bibliographies by failing to synthesize methodologies with their strategic intent. This survey fills this critical gap by employing a systematic literature review (SLR) methodology following PRISMA guidelines and introducing a novel strategic taxonomy centered on three fundamental pillars: Proactive, Reactive, and Attention-Based Relational paradigms. Moving beyond traditional tool-based categorizations, this analysis provides a comparative synthesis of how state-of-the-art algorithms navigate the “information void” during target invisibility. Specifically, a distinction is made between proactive strategies, which anticipate occlusion through motion-based state estimation; reactive strategies, which focus on robust post-occlusion appearance recovery; and attention-based relational paradigms, which maintain object permanence by modeling global spatial-temporal dependencies. To bridge the gap between theoretical taxonomy and practical application, a qualitative comparative synthesis is provided, mapping these strategies to specific scenarios such as self-occlusion, inter-object interactions, and background-induced disappearances. These methodologies are rigorously evaluated across an extensive array of benchmarks, including GOT-10k, LaSOT, TrackingNet, and specialized multi-modal challenges such as RGBT234 and LasHeR. By synthesizing the inherent trade-offs between these paradigms, a definitive shift toward hybrid, context-aware architectures is identified. Finally, this study articulates the existing research bottlenecks, providing a strategic roadmap for the development of more resilient occlusion-handling systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Occlusion handling algorithms in visual object tracking: a survey

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Kazımhan Peker,
  • D. Turgay Altılar

摘要

Although visual object tracking (VOT) has achieved significant milestones, robust occlusion handling remains a formidable challenge, often acting as the primary catalyst for identity switches and tracking failures. While existing related surveys predominantly focus on broad tracking architectures or general deep representation learning, they often treat occlusion as a secondary component and function as annotated bibliographies by failing to synthesize methodologies with their strategic intent. This survey fills this critical gap by employing a systematic literature review (SLR) methodology following PRISMA guidelines and introducing a novel strategic taxonomy centered on three fundamental pillars: Proactive, Reactive, and Attention-Based Relational paradigms. Moving beyond traditional tool-based categorizations, this analysis provides a comparative synthesis of how state-of-the-art algorithms navigate the “information void” during target invisibility. Specifically, a distinction is made between proactive strategies, which anticipate occlusion through motion-based state estimation; reactive strategies, which focus on robust post-occlusion appearance recovery; and attention-based relational paradigms, which maintain object permanence by modeling global spatial-temporal dependencies. To bridge the gap between theoretical taxonomy and practical application, a qualitative comparative synthesis is provided, mapping these strategies to specific scenarios such as self-occlusion, inter-object interactions, and background-induced disappearances. These methodologies are rigorously evaluated across an extensive array of benchmarks, including GOT-10k, LaSOT, TrackingNet, and specialized multi-modal challenges such as RGBT234 and LasHeR. By synthesizing the inherent trade-offs between these paradigms, a definitive shift toward hybrid, context-aware architectures is identified. Finally, this study articulates the existing research bottlenecks, providing a strategic roadmap for the development of more resilient occlusion-handling systems.