Scholarlink

<p>Unsupervised multi-view person re-identification (re-ID) remains challenging in leisure sports environments due to ambient variability and lack of labelled data. This study aimed to develop and validate a robust framework for identity tracking without supervision. Data were collected from 96 participants (48 male, 48 female) across three public leisure sites using 12 fixed cameras capturing 28,800 frames. The proposed AmbientCalib-Net was benchmarked against VITE, SAR, and ACO baselines. Models were trained using ambient-aware calibration, drift-pruning, and multi-level pseudo-labelling. Analyses included embedding similarity, cluster purity, failure mode breakdowns, sensitivity to illumination, activity, and robust paired comparisons. AmbientCalib-Net achieved a mean average precision (mAP) of 0.821 ± 0.037, surpassing VITE (0.694), SAR (0.739), and ACO (0.702). Temporal consistency reached 91.5 ± 3.7%, cluster purity 93.2 ± 4.1%, and assignment accuracy 91.6 ± 3.6%. Occlusion-robustness was 87.9 ± 5.2%, with the lowest failure rate among models (28.1% vs. 43.2–46.2%). Performance declined under high-entropy, low-luminance, and motion-intensive conditions but retained statistical significance in all main metrics (all p &lt; 0.02). AmbientCalib-Net enables accurate, annotation-free identity tracking in complex, real-world sports scenarios.</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

Unsupervised research on multi-view similarity aggregation and multi-level gap optimisation in the re-identification of participants in leisure sports activities

  • 期刊论文

  • 发表日期：

    2026年4月20日

查看全文
  • Hu Jiamin

摘要

Unsupervised multi-view person re-identification (re-ID) remains challenging in leisure sports environments due to ambient variability and lack of labelled data. This study aimed to develop and validate a robust framework for identity tracking without supervision. Data were collected from 96 participants (48 male, 48 female) across three public leisure sites using 12 fixed cameras capturing 28,800 frames. The proposed AmbientCalib-Net was benchmarked against VITE, SAR, and ACO baselines. Models were trained using ambient-aware calibration, drift-pruning, and multi-level pseudo-labelling. Analyses included embedding similarity, cluster purity, failure mode breakdowns, sensitivity to illumination, activity, and robust paired comparisons. AmbientCalib-Net achieved a mean average precision (mAP) of 0.821 ± 0.037, surpassing VITE (0.694), SAR (0.739), and ACO (0.702). Temporal consistency reached 91.5 ± 3.7%, cluster purity 93.2 ± 4.1%, and assignment accuracy 91.6 ± 3.6%. Occlusion-robustness was 87.9 ± 5.2%, with the lowest failure rate among models (28.1% vs. 43.2–46.2%). Performance declined under high-entropy, low-luminance, and motion-intensive conditions but retained statistical significance in all main metrics (all p < 0.02). AmbientCalib-Net enables accurate, annotation-free identity tracking in complex, real-world sports scenarios.