Scholarlink

<p>Motion prediction is a crucial component of multi-object tracking algorithms. Its role is to map the positional information of targets at different time instances to the current moment, establishing correlations in the spatial domain among targets observed at different time points. However, we contend that the estimated target trajectory information in current multi-object tracking algorithms becomes unreliable when targets are obscured and undergo deformation, especially in dim and crowded scenarios. In response to this challenge, we propose the Multi-Head Trajectory Prediction (MHTP) module, a Kalman filtering-based module tailored for lost tracks. By integrating historical motion information, the MHTP module predicts potential track positions in subsequent association processes, enhancing the robustness of track position associations. Additionally, we introduce a distance metric, Multiple Appearance Distance (MAD), combining multi-head predicted positions with appearance information. Leveraging MHTP and MAD, we enhance the data association framework, strengthening the robustness of data associations in target tracking. The approach achieves high accuracy on the test sets of the multi-object tracking datasets MOT17 (HOTA: 66.0%) and MOT20 (HOTA: 63.9%), surpassing the recent state-of-the-art methods.</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 robust approach to deformed pedestrian tracking with multi-trajectory prediction

  • 期刊论文

  • 发表日期：

    2025年4月28日

查看全文
  • Zeqing Wan,
  • Wei Wu

摘要

Motion prediction is a crucial component of multi-object tracking algorithms. Its role is to map the positional information of targets at different time instances to the current moment, establishing correlations in the spatial domain among targets observed at different time points. However, we contend that the estimated target trajectory information in current multi-object tracking algorithms becomes unreliable when targets are obscured and undergo deformation, especially in dim and crowded scenarios. In response to this challenge, we propose the Multi-Head Trajectory Prediction (MHTP) module, a Kalman filtering-based module tailored for lost tracks. By integrating historical motion information, the MHTP module predicts potential track positions in subsequent association processes, enhancing the robustness of track position associations. Additionally, we introduce a distance metric, Multiple Appearance Distance (MAD), combining multi-head predicted positions with appearance information. Leveraging MHTP and MAD, we enhance the data association framework, strengthening the robustness of data associations in target tracking. The approach achieves high accuracy on the test sets of the multi-object tracking datasets MOT17 (HOTA: 66.0%) and MOT20 (HOTA: 63.9%), surpassing the recent state-of-the-art methods.