Scholarlink

<p>Multi-Object Tracking (MOT) remains challenging, especially in complex scenarios with visually similar targets, frequent interactions, and abrupt movements. Existing methods typically emphasize either high accuracy through sophisticated transformer architectures or real-time efficiency via end-to-end frameworks, yet struggle to achieve both simultaneously. Base on this, we introduce a simple and effective method based on the ByteTrack framework. Three improvements are made tailored for complex scenes. Firstly, we propose a small-scale feature extractor enhanced by exponential moving averages (EMA) to capture fine-grained appearance details. Secondly, we employ Enhanced Correlation Coefficient (ECC)-based nonlinear motion compensation to mitigate abrupt positional shifts. Thirdly, we integrate appearance-motion information into a unified cost matrix for robust identity matching. Extensive experiments on MOT20 and SportsMOT datasets validate that our method achieves competitive performance while ensuring real-time inference, achieving a DetA accuracy of 59.5%, an AssA accuracy of 61.1%, and an overall tracking accuracy (HOTA) of 61.4%.</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 real-time multi-object tracking: a strategy in complex scenarios

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Lingqi Xiang,
  • Kexue Sun,
  • Haowei Xie,
  • Xin Tang

摘要

Multi-Object Tracking (MOT) remains challenging, especially in complex scenarios with visually similar targets, frequent interactions, and abrupt movements. Existing methods typically emphasize either high accuracy through sophisticated transformer architectures or real-time efficiency via end-to-end frameworks, yet struggle to achieve both simultaneously. Base on this, we introduce a simple and effective method based on the ByteTrack framework. Three improvements are made tailored for complex scenes. Firstly, we propose a small-scale feature extractor enhanced by exponential moving averages (EMA) to capture fine-grained appearance details. Secondly, we employ Enhanced Correlation Coefficient (ECC)-based nonlinear motion compensation to mitigate abrupt positional shifts. Thirdly, we integrate appearance-motion information into a unified cost matrix for robust identity matching. Extensive experiments on MOT20 and SportsMOT datasets validate that our method achieves competitive performance while ensuring real-time inference, achieving a DetA accuracy of 59.5%, an AssA accuracy of 61.1%, and an overall tracking accuracy (HOTA) of 61.4%.