Advancing UAV Landing Precision: A Comparative Study of Deep Learning Classifiers for Human Detection

This study introduces and implements the use of evidence-based deep learning and Dirichlet distribution for optimal mobility and landing of unmanned aerial vehicles (UAVs). A new evidence-based deep learning model is developed by defining a new loss function on multiple datasets of aerial images captured by drones. The purpose is to identify and process the presence or absence of humans in the landing environment. In this paper we propose a novel deep learning architecture that is capable of providing uncertainty in classification and that can be deployed for input samples. Our project and implemented model are compared with other existing and robust models such as EfficientNET, DenseNet, MobileNet, VGG 16 and 19, Yolov8 and Resnet50. The results show that the use of Evidential Deep Learning (EDL) shows higher accuracy in detecting human presence or absence compared to other existing models. We also model and assess uncertainty for all models used. The use of different models shows the inverse relationship between reliability and uncertainty in person detection in UAVs (higher reliability at the same time as lower uncertainty and vice versa).