Comparison of Drone Classification Accuracy for CNN Models Using UWB Radar

We have previously proposed a method for recognizing various drones from images of ultra-wideband (UWB) radar range profile by using a convolutional neural network (CNN). In general, the classification accuracy for CNN tends to improve as the layer becomes deeper. Therefore, in this paper, we examine the classification accuracy of drones for four types of CNN models (AlexNet, GoogleNet, ResNet-50, and ResNet-101) with different depths. We have investigated the classification performances for five types of drones (Matrice 600, 3DR Solo, Phantom 3, Mavic pro, and Bebop drone) and a radio-controlled flapping bird (Bionic bird) by each of the CNN models. As a result, we have confirmed that all the CNN models can recognize each target with an average accuracy of 90% or more. Furthermore, we have also confirmed that ResNet-50 consisting of 50 layers can recognize all the targets with equal or better accuracy than that by AlexNet which consists of 8 layers.