Scholarlink

Public safety in metropolitan cities is becoming an issue owing to the rising rate of violent crimes, road accidents, and fire accidents. Human observation-based traditional surveillance systems have limited scope to react in real-time. Violence Detection and Alert System (VDAS), a deep learning framework that uses YOLOv5 for object detection, MobileNetV2 for action classification, and ConvLSTM for temporal processing to analyze real-time CCTV footage. When an event takes place, VDAS generates GPS-tagged alerts, wirelessly sent to emergency responders for prompt action. The presented hybrid architecture attains 95% class accuracy, outperforming individual models, at 30 FPS (latency 33 ms/frm) inference rate to assist real-time detection. The system is verified using precision (96.2%), recall (94.8%), and F1-score (95.5%), proving the efficiency of violence, accidents, and fire risk. Error analysis demonstrates misclassification mainly takes place in densely populated areas, which would be addressed in future work. Testing on real-world scenes validates VDAS’s application to urban cities, school campuses, shopping centers, and other public places. In addition, VDAS can be computationally lightweight, with reduced GPU memory and processing required compared to Transformer-based models, making it appropriate for real-time edge deployment. Future improvement entails multi-camera integration, audio violence detection (e.g., gunshot, scream), edge device computational efficiency optimization, and mass-scale real-world deployment to further enhance public safety monitoring.

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

Violence Detection and Alert System (VDAS): A Real-Time Public Safety Solution Using Deep Learning and GPS Integration

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rameswar Balaji,
  • Karthikeya,
  • Mohana Krishna,
  • Nithish,
  • G. Radhika

摘要

Public safety in metropolitan cities is becoming an issue owing to the rising rate of violent crimes, road accidents, and fire accidents. Human observation-based traditional surveillance systems have limited scope to react in real-time. Violence Detection and Alert System (VDAS), a deep learning framework that uses YOLOv5 for object detection, MobileNetV2 for action classification, and ConvLSTM for temporal processing to analyze real-time CCTV footage. When an event takes place, VDAS generates GPS-tagged alerts, wirelessly sent to emergency responders for prompt action. The presented hybrid architecture attains 95% class accuracy, outperforming individual models, at 30 FPS (latency 33 ms/frm) inference rate to assist real-time detection. The system is verified using precision (96.2%), recall (94.8%), and F1-score (95.5%), proving the efficiency of violence, accidents, and fire risk. Error analysis demonstrates misclassification mainly takes place in densely populated areas, which would be addressed in future work. Testing on real-world scenes validates VDAS’s application to urban cities, school campuses, shopping centers, and other public places. In addition, VDAS can be computationally lightweight, with reduced GPU memory and processing required compared to Transformer-based models, making it appropriate for real-time edge deployment. Future improvement entails multi-camera integration, audio violence detection (e.g., gunshot, scream), edge device computational efficiency optimization, and mass-scale real-world deployment to further enhance public safety monitoring.