Drone Model Identification by Convolutional Neural Network from Video Stream

Abstract

We present a convolutional neural network model that correctly identifies drone models in real-life video streams of flying drones. To achieve this, we show a method of generating synthetic drone images. To create a diverse dataset, the simulation parameters (such as drone textures, lighting, and orientation) are randomized. This synthetic dataset is used to train a convolutional neural network to identify the drone model: DJI Phantom, DJI Mavic, or DJI Inspire. The model is then tested on a real-life Anti-UAV dataset of flying drones. The benchmark results show that the DenseNet201 architecture performed the best. Adding Gaussian noise to the training dataset and performing full training (as opposed to freezing layers) shows the best results. The model shows an average accuracy of 92.4%, and an average precision of 88.6% on the test dataset.

Drone Model Classification Using Convolutional Neural Network Trained on Synthetic Data

“…Certain artefacts, such as white and black lines, also make the classification task challenging. Extracts from the dataset can be seen in [ 1 , 20 ].…”

“…This is achieved by creating a synthetic dataset to train the neural network and testing its performance on a real-life dataset of flying drones. It is an extension study of [ 1 ] that examines more closely the effects of synthetic noise, dataset size, and simulation parameters (ablation study). It also updates on the latest literature in the related field of drone detection.…”