Scholarlink

<p>Multi-object tracking (MOT) in surveillance scenarios demands both high accuracy and computational efficiency, especially under challenging conditions such as occlusion, scale variation, and dense targets. While one-stage joint detection and embedding models offer promising real-time performance, they often suffer from task inconsistency and feature scale mismatch between detection and re-identification (ReID) branches, leading to degraded appearance embeddings and frequent identity switches. To address this issue, this paper proposes a novel joint detection and appearance embedding based on YOLOv9: YOLO-JDE. We use a kind of dynamic-aware downsampling (DAD) in the original model to enhance the network's capability to extract global feature information, which effectively improves detection accuracy. Due to the intrinsic dissimilarity between the detection task and the re-identification (ReID) task, we introduce a multi-task decomposition network (MTDNet) that mitigates the conflict between these two tasks by learning the dissimilarity to output different feature maps. In addition, we propose a multi-scale aggregation network (MSANet) in the ReID branch to address the issue of identity switches caused by occlusion, scale variation, and dense targets during the tracking and matching process. In the final stage of data association, we propose a multilevel association matching approach that leverages confidence, appearance, and IoU distance, which can achieve about 0.7% performance improvement on multiple tracking models. Through validation on the public MOT16, MOT17, and MOT20 test sets for surveillance scenarios, YOLO-JDE achieved MOTA metrics of 78.2%, 77.4%, and 73.0%, respectively.</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

Multi-object tracking in surveillance scenarios based on joint detection and multi-feature fusion

  • 期刊论文

  • 发表日期：

    2025年5月31日

查看全文
  • Runxing Zhao,
  • Jianxin Gong,
  • Zhiwen Wang

摘要

Multi-object tracking (MOT) in surveillance scenarios demands both high accuracy and computational efficiency, especially under challenging conditions such as occlusion, scale variation, and dense targets. While one-stage joint detection and embedding models offer promising real-time performance, they often suffer from task inconsistency and feature scale mismatch between detection and re-identification (ReID) branches, leading to degraded appearance embeddings and frequent identity switches. To address this issue, this paper proposes a novel joint detection and appearance embedding based on YOLOv9: YOLO-JDE. We use a kind of dynamic-aware downsampling (DAD) in the original model to enhance the network's capability to extract global feature information, which effectively improves detection accuracy. Due to the intrinsic dissimilarity between the detection task and the re-identification (ReID) task, we introduce a multi-task decomposition network (MTDNet) that mitigates the conflict between these two tasks by learning the dissimilarity to output different feature maps. In addition, we propose a multi-scale aggregation network (MSANet) in the ReID branch to address the issue of identity switches caused by occlusion, scale variation, and dense targets during the tracking and matching process. In the final stage of data association, we propose a multilevel association matching approach that leverages confidence, appearance, and IoU distance, which can achieve about 0.7% performance improvement on multiple tracking models. Through validation on the public MOT16, MOT17, and MOT20 test sets for surveillance scenarios, YOLO-JDE achieved MOTA metrics of 78.2%, 77.4%, and 73.0%, respectively.