Scholarlink

<p>Locating and tracking a specific person of interest in a single visual query remains a significant challenge in complex surveillance environments. Current paradigms fall short: generic multi-object trackers suffer from identity loss over time, while existing person search methods, designed for static image galleries, lack robustness against the dynamic complexities of video streams, especially occlusions. This paper introduces QueryTrack, a comprehensive framework designed specifically for this query-based tracking task. The core novelty lies in a powerful re-identification engine that fuses four distinct feature types—HOG, Gabor, Color, and VGG16—into a highly discriminative signature for the target. This signature drives a hybrid tracking algorithm that synergizes motion prediction and visual tracking to maintain identity continuity. Furthermore, we propose a new post-occlusion recovery technique to handle long-term disappearances. Experimental evaluations validate our method’s superior performance, achieving F1 scores of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(97.20\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>97.20</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> in crowded scenarios and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(96.35\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>96.35</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> with minimal occlusion, confirming its significant contribution to accurate and persistent person tracking under realistic conditions. Additionally, we provide a transparent computational cost analysis, confirming the system’s viability for offline forensic investigation where accuracy is paramount.</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

QueryTrack: identifying and tracking a person of interest using clothing-based hybrid features

  • 期刊论文

  • 发表日期：

    2026年2月1日

查看全文
  • Gizem Ortac Kosun,
  • Seckin Yilmaz,
  • Ruya Samli

摘要

Locating and tracking a specific person of interest in a single visual query remains a significant challenge in complex surveillance environments. Current paradigms fall short: generic multi-object trackers suffer from identity loss over time, while existing person search methods, designed for static image galleries, lack robustness against the dynamic complexities of video streams, especially occlusions. This paper introduces QueryTrack, a comprehensive framework designed specifically for this query-based tracking task. The core novelty lies in a powerful re-identification engine that fuses four distinct feature types—HOG, Gabor, Color, and VGG16—into a highly discriminative signature for the target. This signature drives a hybrid tracking algorithm that synergizes motion prediction and visual tracking to maintain identity continuity. Furthermore, we propose a new post-occlusion recovery technique to handle long-term disappearances. Experimental evaluations validate our method’s superior performance, achieving F1 scores of \(97.20\%\) 97.20 % in crowded scenarios and \(96.35\%\) 96.35 % with minimal occlusion, confirming its significant contribution to accurate and persistent person tracking under realistic conditions. Additionally, we provide a transparent computational cost analysis, confirming the system’s viability for offline forensic investigation where accuracy is paramount.