Human Detection in UAV Thermal Imagery: Dataset Extension and Comparative Evaluation on Embedded Platforms.

Area of Science:

• Computer Vision
• Robotics
• Remote Sensing

Background:

• Unmanned aerial vehicles (UAVs) with thermal cameras are vital for search and rescue (SAR), but existing datasets lack mountainous terrain data, hindering model performance.
• Detection challenges in SAR include low visibility and small human targets, especially in complex environments like mountains.

Purpose of the Study:

• To address the limitations of existing datasets by creating a new, extensive benchmark of thermal images from mountainous environments.
• To conduct a comprehensive comparative evaluation of object detection and semantic segmentation models for human detection in SAR operations on embedded hardware.

Main Methods:

• Collected and annotated a new dataset of thermal images in mountainous regions using a DJI M3T drone.
• Integrated the new dataset with ten existing sources to create a benchmark of over 75,000 images.
• Evaluated object detection (YOLOv8/9/10, RT-DETR) and semantic segmentation (U-Net variants) models on an NVIDIA Jetson AGX Orin, analyzing accuracy, speed, and energy consumption.

Main Results:

• Both semantic segmentation and object detection models achieved high accuracy for human detection in mountain scenarios (90% and 85% respectively).
• Segmentation models achieved real-time performance (up to 27 FPS in FP16 mode) on the embedded Jetson platform.
• Comparative analysis provided insights into the trade-offs between different models regarding accuracy, inference speed, and energy efficiency.

Conclusions:

• The developed mountainous thermal image dataset significantly enhances benchmarks for SAR applications.
• Semantic segmentation and object detection models are effective for human detection in challenging mountain environments using UAVs.
• The comprehensive evaluation provides valuable data for deploying efficient AI models on embedded hardware for real-world SAR missions.