Fast classification of drones and birds with an LSTM network applied to 1D phase data

Abstract

This study investigates a new type of drone classifier based on Long Short-Term Memory (LSTM) networks. As a real-time surveillance system, the classification time of a drone detection radar is crucial. The motivation for this work is to develop a classification framework which has low latency in terms of data processing for the algorithm input. Theoretical modeling of a rotary wing drone and a bird wing flapping returns were done first to exhibit the difference in the patterns of the respective phase progressions. Then, 94 GHzexperimental trial data containing 4800 sequences of drones, birds, noise and clutter were used to create a diverse training dataset of 1D phase data for supervised learning. A stackedLSTM network with tuned hyperparameters was generated to reduce the possible overfitting from a simple LSTM model. Validation accuracy of 98.1% was achieved for 2-class classification of drone and non-drone. Further performance assessment was then done with 30 unseen test data, where the network was able to correctly classify all the sequences. It is ascertained that this method can be ~10 times faster than a spectrogram based classification model, which requires additional Fast Fourier Transform (FFT) operations.