Scholarlink

<p>To address the challenges of inadequate fusion and redundancy in the interaction between template and search features within visual encoders, we propose an Asynchronous Fusion Tracking Model based on Visual Receptive Weighted Key-Value Perceptual(VAFTrack). Initially, we introduce a Visual Receptive Weighted Key-Value Perceptual Fusion Module(VPFM) that enhances the matching accuracy between template and search features by sequentially integrating Contextual Spatial Perception Module(CSPM) and Contextual Channel Perception Module(CCPM). This approach improves the model’s target awareness and adaptability to variations in scenes. Subsequently, we design a Spatial Optimization Enhancement Module(SOEM) and a Channel Optimization Enhancement Module(COEM) to mitigate redundancy and augment information density by optimizing spatial and channel representations within feature maps. These enhancement bolsters the model’s robustness against occlusion and complex backgrounds, thereby strengthening its discriminative capacity for target tracking. The proposed model achieves state-of-the-art performance, with accuracies of 81.9% on TrackingNet and 73.6% on LaSOT, and an overlap rate of 73.6% on the GOT-10k dataset.</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

VAFTrack: asynchronous feature fusion via visual receptive weighted key-value perceptual for visual tracking

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Zhongmin Liu,
  • Zhenhua Li,
  • Wenjin Hu

摘要

To address the challenges of inadequate fusion and redundancy in the interaction between template and search features within visual encoders, we propose an Asynchronous Fusion Tracking Model based on Visual Receptive Weighted Key-Value Perceptual(VAFTrack). Initially, we introduce a Visual Receptive Weighted Key-Value Perceptual Fusion Module(VPFM) that enhances the matching accuracy between template and search features by sequentially integrating Contextual Spatial Perception Module(CSPM) and Contextual Channel Perception Module(CCPM). This approach improves the model’s target awareness and adaptability to variations in scenes. Subsequently, we design a Spatial Optimization Enhancement Module(SOEM) and a Channel Optimization Enhancement Module(COEM) to mitigate redundancy and augment information density by optimizing spatial and channel representations within feature maps. These enhancement bolsters the model’s robustness against occlusion and complex backgrounds, thereby strengthening its discriminative capacity for target tracking. The proposed model achieves state-of-the-art performance, with accuracies of 81.9% on TrackingNet and 73.6% on LaSOT, and an overlap rate of 73.6% on the GOT-10k dataset.