Deep Convolutional Neural Networks for Real-Time Human Detection and Tracking on UAVs Embedded Systems

Human detection for critical missions with unmanned aerial vehicle (UAV) support becomes more and more important in the actual context when tension at borders builds up for an increasing number of countries. Although convolutional neural networks are continuously evolving, the required computational resources pose a great problem when implemented on portable embedded systems such as UAVs, with limited processing power and autonomy. This demand becomes even more drastic when running real-time human detection and tracking. This paper proposes an improved implementation of the YOLOv.7, trained on a custom dataset, for real-time human detection and tracking with confidence scores above 80% on NVIDIA Jetson TX2 neural processing unit equipped on DJI Matrice 100 UAV. The authors created a YOLOv.7 model running independently on an embedded system for real-time human detection and tracking.