Scholarlink

<p>Siamese-based trackers achieve much progress in accuracy and tracking speeds, which use the cross-correlation to compute the target similarity. However, these trackers based on cross-correlation ignore the spatial layout of feature maps and the correspondences of the features between the template and search regions. And these trackers based on cross-correlation either lose lots of foreground information or retain numerous background information due to pre-fixed feature regions. We design a Kronecker Product Matching based feature fusion network for establishing the spatial region correspondences from the target templates to search images. The spatial information of template features and search region features are obtained, which is helpful to obtain more accurate target similarity. In addition, a normalization based attention module is introduced in the template branch to suppress less salient features and background information. By integrating the designed feature fusion network and the attention module, a novel tracking algorithm is proposed in the Siamese tracking framework. Extensive experiments on six challenging benchmarks including OTB-100, GOT-10k, LaSOT, UAV123, VOT2018 and NFS demonstrate the generalization ability and effectiveness of the proposed tracker. In particular, the proposed tracker achieves the AUC score of 63.7% on GOT-10k, 62.6% on UAV123 and precision score of 89.6% on OTB-100, while running at 40 frames per second (FPS).</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

Spatial correspondence matching based feature fusion for object tracking with attention learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月6日

查看全文
  • Yuanyun Wang,
  • Wenhui Yang,
  • Lingtao Zhou,
  • Shenmiao Jin,
  • Jun Wang

摘要

Siamese-based trackers achieve much progress in accuracy and tracking speeds, which use the cross-correlation to compute the target similarity. However, these trackers based on cross-correlation ignore the spatial layout of feature maps and the correspondences of the features between the template and search regions. And these trackers based on cross-correlation either lose lots of foreground information or retain numerous background information due to pre-fixed feature regions. We design a Kronecker Product Matching based feature fusion network for establishing the spatial region correspondences from the target templates to search images. The spatial information of template features and search region features are obtained, which is helpful to obtain more accurate target similarity. In addition, a normalization based attention module is introduced in the template branch to suppress less salient features and background information. By integrating the designed feature fusion network and the attention module, a novel tracking algorithm is proposed in the Siamese tracking framework. Extensive experiments on six challenging benchmarks including OTB-100, GOT-10k, LaSOT, UAV123, VOT2018 and NFS demonstrate the generalization ability and effectiveness of the proposed tracker. In particular, the proposed tracker achieves the AUC score of 63.7% on GOT-10k, 62.6% on UAV123 and precision score of 89.6% on OTB-100, while running at 40 frames per second (FPS).