Scholarlink

<p>Object tracking faces several critical challenges, including model redundancy, insufficient feature representation for small objects, difficulties in modeling nonlinear motion, and Identity Switches (IDs). To address these issues, this paper proposes a lightweight and robust pedestrian small object tracking algorithm named PerTrack on the basis of the ByteTrack framework. The core innovations are as follows: a Dynamic Weight Allocation Dilated Convolution (DWA-Conv) unit is adopted to strengthen multi-scale feature representation for small objects; a Dual-Factor Adaptive Kalman Filter (AKF) mechanism is designed to address the limitations of traditional linear motion models; a Scene-aware Multi-modal Hierarchical Data Association (SDA) strategy is constructed to improve identity preservation in occluded and crowded scenes. Experimental results on MOT17, MOT20, and VisDrone datasets demonstrate that compared with ByteTrack, PerTrack achieves 5.9% improvement in Multiple Object Tracking Accuracy (MOTA) and 6.9% improvement in Identity F1-score (IDF1) on MOT17, with 203 fewer IDs and only 4.87&#xa0;M parameters. It also achieves state-of-the-art performance on MOT20 and VisDrone in small object scenarios. The proposed algorithm achieves higher tracking accuracy and identity consistency under a lightweight architecture, which is suitable for deployment on mobile and embedded edge devices.</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

A lightweight pedestrian small object tracking algorithm pertrack for edge deployment

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月19日

查看全文
  • Yuan Wang,
  • Peizhen Xing,
  • Xiaoshu Hu,
  • Jie Shao

摘要

Object tracking faces several critical challenges, including model redundancy, insufficient feature representation for small objects, difficulties in modeling nonlinear motion, and Identity Switches (IDs). To address these issues, this paper proposes a lightweight and robust pedestrian small object tracking algorithm named PerTrack on the basis of the ByteTrack framework. The core innovations are as follows: a Dynamic Weight Allocation Dilated Convolution (DWA-Conv) unit is adopted to strengthen multi-scale feature representation for small objects; a Dual-Factor Adaptive Kalman Filter (AKF) mechanism is designed to address the limitations of traditional linear motion models; a Scene-aware Multi-modal Hierarchical Data Association (SDA) strategy is constructed to improve identity preservation in occluded and crowded scenes. Experimental results on MOT17, MOT20, and VisDrone datasets demonstrate that compared with ByteTrack, PerTrack achieves 5.9% improvement in Multiple Object Tracking Accuracy (MOTA) and 6.9% improvement in Identity F1-score (IDF1) on MOT17, with 203 fewer IDs and only 4.87 M parameters. It also achieves state-of-the-art performance on MOT20 and VisDrone in small object scenarios. The proposed algorithm achieves higher tracking accuracy and identity consistency under a lightweight architecture, which is suitable for deployment on mobile and embedded edge devices.