Scholarlink

Avalanches pose a significant threat in mountainous regions, necessitating efficient and real-time victim detection for effective search and rescue operations. Traditional methods often face limitations due to poor visibility, challenging terrain, and dependence on beacon devices, which may not always be carried by victims. This research introduces a UAV-based system that integrates deep learning techniques to improve the detection of avalanche victims. The system utilizes a quadcopter equipped with high-resolution cameras, GPS modules, and on-board flight data processing units. Initially, a hybrid InceptionV3-SVM model was employed, achieving an accuracy of 0.86. To improve real-time performance, a comparative study was conducted on various YOLO models, and YOLOv11n was later adopted along with a custom dataset, offering reduced latency while maintaining detection reliability. The system incorporates Zero-Reference Deep Curve Estimation (Zero-DCE) to enhance visibility in low-light conditions and a Hidden Markov Model (HMM) for temporal consistency, reducing false positives. The dataset for training includes custom avalanche scenarios and images from the Open Images Dataset, focusing on various human body parts for comprehensive detection capabilities.

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

UAV-Based Avalanche Victim Detection: An Integrated Approach Using Deep Learning and Aerial Surveillance

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jhalak Dutta,
  • Smita Das,
  • Soumyadeep Bose,
  • Monojit Das,
  • Sampriti Mitra,
  • Sourjya Mukherjee,
  • Srijita Chatterjee,
  • Debanjan Sahana

摘要

Avalanches pose a significant threat in mountainous regions, necessitating efficient and real-time victim detection for effective search and rescue operations. Traditional methods often face limitations due to poor visibility, challenging terrain, and dependence on beacon devices, which may not always be carried by victims. This research introduces a UAV-based system that integrates deep learning techniques to improve the detection of avalanche victims. The system utilizes a quadcopter equipped with high-resolution cameras, GPS modules, and on-board flight data processing units. Initially, a hybrid InceptionV3-SVM model was employed, achieving an accuracy of 0.86. To improve real-time performance, a comparative study was conducted on various YOLO models, and YOLOv11n was later adopted along with a custom dataset, offering reduced latency while maintaining detection reliability. The system incorporates Zero-Reference Deep Curve Estimation (Zero-DCE) to enhance visibility in low-light conditions and a Hidden Markov Model (HMM) for temporal consistency, reducing false positives. The dataset for training includes custom avalanche scenarios and images from the Open Images Dataset, focusing on various human body parts for comprehensive detection capabilities.