Human Detection in Restricted Areas Using Deep Convolutional Neural Networks

With the help of state-of-the-art DCNNs precise detection of persons is possible in images and videos acquired from UAVs at low and medium altitudes. The current paper proposes for comparison of two DCNNs: Scaled-YOLOv4 and YOLOv7 trained on a custom dataset through transfer learning with data acquired from UAV. The aim is to take advantage of the capabilities of YOLOv7 to train a lightweight model with good accuracy that can be loaded on embedded systems present onboard UAVs capable of real-time person detection. Through transfer learning, the model achieves detection scores above 90% at altitudes of 30m using YOLOv7 architecture. Experiments were conducted to prove its ability to successfully multiple human detection frame-by-frame with over 70% confidence scores.