Scholarlink

<p>Visual tracking in dynamic real-world environments is challenging due to rapid motion, occlusion, cluttered backgrounds, and the visual similarity among objects. Conventional memory-based trackers partially alleviate these problems, but appearance memory is prone to drift under fast dynamics, while semantic memory often suffers from redundancy and ambiguity when targets blend into complex scenes. This work introduces Episodic–Semantic Memory Tracking (ESMT), a unified framework that integrates temporal and semantic modules for robust vision–language tracking. Specifically, the Temporal-State Episodic Memory preserves motion-consistent states through temporal state updating and selective retrieval, while the Graph-Guided Semantic Memory enhances visual features by constructing a semantic similarity graph and aligning it with linguistic cues, jointly suppressing redundant background responses and strengthening target discrimination. Furthermore, a Context-Adaptive Fusion Gate balances episodic and semantic contributions according to scene dynamics. Experiments on LaSOT, TNL2K, and OTB99-L confirm that ESMT achieves competitive accuracy and robustness, demonstrating consistent performance across diverse and visually complex tracking scenarios.</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

ESMT: Context-adaptive vision-language tracking with episodic-semantic memory

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Jiapeng Zhang,
  • Ying Wei,
  • Yongfeng Li,
  • Gang Yang

摘要

Visual tracking in dynamic real-world environments is challenging due to rapid motion, occlusion, cluttered backgrounds, and the visual similarity among objects. Conventional memory-based trackers partially alleviate these problems, but appearance memory is prone to drift under fast dynamics, while semantic memory often suffers from redundancy and ambiguity when targets blend into complex scenes. This work introduces Episodic–Semantic Memory Tracking (ESMT), a unified framework that integrates temporal and semantic modules for robust vision–language tracking. Specifically, the Temporal-State Episodic Memory preserves motion-consistent states through temporal state updating and selective retrieval, while the Graph-Guided Semantic Memory enhances visual features by constructing a semantic similarity graph and aligning it with linguistic cues, jointly suppressing redundant background responses and strengthening target discrimination. Furthermore, a Context-Adaptive Fusion Gate balances episodic and semantic contributions according to scene dynamics. Experiments on LaSOT, TNL2K, and OTB99-L confirm that ESMT achieves competitive accuracy and robustness, demonstrating consistent performance across diverse and visually complex tracking scenarios.