Scholarlink

<p>Multi-object tracking (MOT) aims to estimate the bounding box and identity of objects in a video. Existing MOT methods often struggle with target occlusion and the under-utilization of detection information, leading to fragmented trajectories and failure to promptly track emerging targets. In this paper, we propose a robust and adaptive multi-target tracking framework. It introduces an adaptive thresholding strategy in global greedy matching, which dynamically adjusts to the occlusion scenario and improves flexibility and accuracy. Secondly, it utilizes the extended trajectory method to associate the remaining and emerging target trajectories with the unmatched high-scoring detected targets, which enhances the utilization of detection information and tracking continuity. Finally, to further refine prediction accuracy, the framework uses optimized Kalman filtering and camera motion compensation. An improved Gaussian interpolation method is also introduced to smooth the tracking trajectory, addressing challenges such as long-term occlusion and nonlinear motion. Extensive experiments on the MOT17 and MOT20 datasets demonstrate the superiority of our approach, achieving competitive performance in MOTA, IDF1, and HOTA. The source code and pre-trained models with deployable versions are available at <a href="https://github.com/wwwshengbo/ROT-SORT">https://github.com/wwwshengbo/ROT-SORT</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

Adaptive and extended trajectory matching for robust multi-target tracking

  • 期刊论文

  • 发表日期：

    2025年4月9日

查看全文
  • Shengbo Wang,
  • Xuedong Guo,
  • Lei Deng,
  • Guangkai Sun,
  • Mingli Dong

摘要

Multi-object tracking (MOT) aims to estimate the bounding box and identity of objects in a video. Existing MOT methods often struggle with target occlusion and the under-utilization of detection information, leading to fragmented trajectories and failure to promptly track emerging targets. In this paper, we propose a robust and adaptive multi-target tracking framework. It introduces an adaptive thresholding strategy in global greedy matching, which dynamically adjusts to the occlusion scenario and improves flexibility and accuracy. Secondly, it utilizes the extended trajectory method to associate the remaining and emerging target trajectories with the unmatched high-scoring detected targets, which enhances the utilization of detection information and tracking continuity. Finally, to further refine prediction accuracy, the framework uses optimized Kalman filtering and camera motion compensation. An improved Gaussian interpolation method is also introduced to smooth the tracking trajectory, addressing challenges such as long-term occlusion and nonlinear motion. Extensive experiments on the MOT17 and MOT20 datasets demonstrate the superiority of our approach, achieving competitive performance in MOTA, IDF1, and HOTA. The source code and pre-trained models with deployable versions are available at https://github.com/wwwshengbo/ROT-SORT.