Scholarlink

<p>Detecting violent actions in extremely low-resolution (eLR) surveillance footage is challenging due to severe pixel degradation and motion ambiguity. This paper presents GAN-PPP, an integrated framework combining YOLOv8-based eLR object detection, domain-adapted SRGANs, and Mixture-of-Experts (MoE) LSTM modeling for enhanced violence recognition. The system detects action-relevant regions at 12<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>16 resolution, applies specialized SRGANs to recover fine pedestrian and pose details, and analyzes the restored sequences via MoE-LSTM to capture critical temporal dynamics. Experiments on HMDB51, UCF101, KTH, UCSD, Avenue, and Collective Activity (in eLR form) show that GAN-PPP improves accuracy by 11.4% over baseline eLR methods and surpasses C3D and LRCN by 9.2% and 7.8% on HMDB51-eLR. On UCF101-eLR, it achieves a 6.7% F1-score gain, confirming the value of super-resolution and expert-guided temporal fusion. GAN-PPP provides an effective ROI-based eLR violence-recognition framework and demonstrates practical inference speed on RTX 3090 hardware; however, deployment-level real-time performance under heterogeneous edge and surveillance conditions requires further validation.</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

GAN-PPP: super-resolution GAN-based pedestrian pose pinpointing for enhanced violence detection

  • 期刊论文

  • 发表日期：

    2026年6月27日

查看全文
  • Mahmoud Taha,
  • Ahmed B. Zaky,
  • Hirozumi Yamaguchi,
  • Ahmed Fares

摘要

Detecting violent actions in extremely low-resolution (eLR) surveillance footage is challenging due to severe pixel degradation and motion ambiguity. This paper presents GAN-PPP, an integrated framework combining YOLOv8-based eLR object detection, domain-adapted SRGANs, and Mixture-of-Experts (MoE) LSTM modeling for enhanced violence recognition. The system detects action-relevant regions at 12 \(\times\) 16 resolution, applies specialized SRGANs to recover fine pedestrian and pose details, and analyzes the restored sequences via MoE-LSTM to capture critical temporal dynamics. Experiments on HMDB51, UCF101, KTH, UCSD, Avenue, and Collective Activity (in eLR form) show that GAN-PPP improves accuracy by 11.4% over baseline eLR methods and surpasses C3D and LRCN by 9.2% and 7.8% on HMDB51-eLR. On UCF101-eLR, it achieves a 6.7% F1-score gain, confirming the value of super-resolution and expert-guided temporal fusion. GAN-PPP provides an effective ROI-based eLR violence-recognition framework and demonstrates practical inference speed on RTX 3090 hardware; however, deployment-level real-time performance under heterogeneous edge and surveillance conditions requires further validation.