Scholarlink

<p>This paper presents an efficient, real-time fisheye camera-based framework for wide-area behavior monitoring. Unlike previous methods that rely on single object detection models struggling to generalize across environments due to limited training data, our approach decomposes behavior recognition into general-purpose sub-tasks—object detection, face detection, and facial landmark estimation—each handled by a lightweight model trained on large, publicly available datasets. Rule-based post-processing then combines their outputs to recognize behaviors without requiring domain-specific action labels, enabling strong generalizability across settings. The system detects three key student behaviors: laptop use, walking, and talking, which are indicators of engagement. Fisheye distortion is mitigated through calibration-based rectification, allowing a single camera to monitor a wide area. Computational efficiency is improved via TensorRT conversion and parallel processing, achieving real-time inference at up to 67 frames per second. Evaluated on a custom laboratory dataset with 791 annotated action instances, the system achieves an overall F1 score of 0.824 and substantially outperforms spatio-temporal baselines (YOWOv3: F1 = 0.572), demonstrating the effectiveness of our multi-model decomposition. While demonstrated in educational laboratories, the approach is scalable and deployable in diverse real-world environments requiring wide-area analytics.</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

Real-time behavior detection in wide-area environments: a lightweight multi-model framework optimized for fisheye camera

  • 期刊论文

  • 发表日期：

    2026年8月2日

查看全文
  • Tuan Linh Dang,
  • Xuan Tung Tran,
  • Huu Phuc Hoang,
  • Nhat Hai Nguyen

摘要

This paper presents an efficient, real-time fisheye camera-based framework for wide-area behavior monitoring. Unlike previous methods that rely on single object detection models struggling to generalize across environments due to limited training data, our approach decomposes behavior recognition into general-purpose sub-tasks—object detection, face detection, and facial landmark estimation—each handled by a lightweight model trained on large, publicly available datasets. Rule-based post-processing then combines their outputs to recognize behaviors without requiring domain-specific action labels, enabling strong generalizability across settings. The system detects three key student behaviors: laptop use, walking, and talking, which are indicators of engagement. Fisheye distortion is mitigated through calibration-based rectification, allowing a single camera to monitor a wide area. Computational efficiency is improved via TensorRT conversion and parallel processing, achieving real-time inference at up to 67 frames per second. Evaluated on a custom laboratory dataset with 791 annotated action instances, the system achieves an overall F1 score of 0.824 and substantially outperforms spatio-temporal baselines (YOWOv3: F1 = 0.572), demonstrating the effectiveness of our multi-model decomposition. While demonstrated in educational laboratories, the approach is scalable and deployable in diverse real-world environments requiring wide-area analytics.