Scholarlink

<p>Point-based 3D single-object tracking (SOT) is an important task in the field of computer vision. The earliest appearance matching methods are sensitive to interference factors due to the characteristics of point clouds. Later motion-centric methods only grasp short-term motion information and ignore long-term motion clues. In order to break the boundary between the two methods, we propose a SOT method that uses feature correlation to correct motion offset. This method calculates the similarity of motion features between consecutive frames to correct the long-term historical motion offset, and then accurately locates a single target. We introduced a BMOP module to predict rough long-term historical motion clues, and also introduced IFE and DMFF modules to interact and fuse continuous frame features at multiple scales to obtain dense motion features. Finally, the MCE module corrects the offset to obtain accurate motion estimation of the final box. Extensive experiments show that our method exhibits excellent robustness and surpasses multiple classic methods. The code will be released at <a href="https://github.com/wangjie16138/PCtrack">https://github.com/wangjie16138/PCtrack</a>.</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

3D Single-Object Tracking Method with Feature-Correlation Motion Offset Correction

  • 期刊论文

  • 发表日期：

    2025年9月16日

查看全文
  • Jie Wang,
  • Yue Yu,
  • Jietao Cheng,
  • Jun Li,
  • Jun Tang

摘要

Point-based 3D single-object tracking (SOT) is an important task in the field of computer vision. The earliest appearance matching methods are sensitive to interference factors due to the characteristics of point clouds. Later motion-centric methods only grasp short-term motion information and ignore long-term motion clues. In order to break the boundary between the two methods, we propose a SOT method that uses feature correlation to correct motion offset. This method calculates the similarity of motion features between consecutive frames to correct the long-term historical motion offset, and then accurately locates a single target. We introduced a BMOP module to predict rough long-term historical motion clues, and also introduced IFE and DMFF modules to interact and fuse continuous frame features at multiple scales to obtain dense motion features. Finally, the MCE module corrects the offset to obtain accurate motion estimation of the final box. Extensive experiments show that our method exhibits excellent robustness and surpasses multiple classic methods. The code will be released at https://github.com/wangjie16138/PCtrack.