Scholarlink

<p>Lightweight target tracking has made significant progress in recent years, but these algorithms face the challenge of balancing accuracy and efficiency. In this work, we propose an efficient lightweight tracking, named&#xa0;ELGSOT, which is based on separable mixed attention and the local–global feature fusion Transformer. The proposed backbone employs wavelet convolution (WTConv) optimization to fuse template and search regions during Transformer-based feature extraction, generating high-quality feature encodings. Our Local–Global Feature Fusion Network (LGFFN) leverages separable mixed attention to effectively integrate local and global features for robust target state estimation and global context modeling. We also simulate complex environments using sampling-based training strategies, enhancing the tracker’s robustness and stability. Our ablation studies demonstrate the effectiveness of combining LGFFN with WTConv. Experimental results show that ELGSOT outperforms related lightweight trackers on five challenging benchmarks (GOT-10k, TrackingNet, LaSOT, UAV123, OTB100), achieving real-time performance (36 <i>fps</i>) on CPU with 3.78 M parameters and 0.23 G FLOPs.</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

Efficient local–global feature fusion transformer for siamese object tracking

  • 期刊论文

  • 发表日期：

    2025年6月10日

查看全文
  • Fei Pan,
  • Lianyu Zhao,
  • Chenglin Wang,
  • Xianfeng Wang,
  • Guangxin Ren

摘要

Lightweight target tracking has made significant progress in recent years, but these algorithms face the challenge of balancing accuracy and efficiency. In this work, we propose an efficient lightweight tracking, named ELGSOT, which is based on separable mixed attention and the local–global feature fusion Transformer. The proposed backbone employs wavelet convolution (WTConv) optimization to fuse template and search regions during Transformer-based feature extraction, generating high-quality feature encodings. Our Local–Global Feature Fusion Network (LGFFN) leverages separable mixed attention to effectively integrate local and global features for robust target state estimation and global context modeling. We also simulate complex environments using sampling-based training strategies, enhancing the tracker’s robustness and stability. Our ablation studies demonstrate the effectiveness of combining LGFFN with WTConv. Experimental results show that ELGSOT outperforms related lightweight trackers on five challenging benchmarks (GOT-10k, TrackingNet, LaSOT, UAV123, OTB100), achieving real-time performance (36 fps) on CPU with 3.78 M parameters and 0.23 G FLOPs.