Urban Bird-Drone Classification With Synthetic Micro-Doppler Spectrograms
Daniel White, Mohammed Jahangir, Chris Baker, Michail Antoniou
Abstract
In this article, a method for creating highly realistic synthetic drone micro-Doppler spectrograms is presented and its effectiveness of training a bird-drone classifier for real scenario classification is shown via comparisons to a real benchmark. The effect of drone motor speed sampling used when simulating drone micro-Doppler is shown to have a significant impact on the accuracy of synthetic results and variations of this approach are explored. Four synthetic datasets were created differing in motor speed sampling and each were compared in their ability to train a convolutional neural network to classify real data. The highest fidelity synthetic dataset achieved a classification accuracy of 86.6% compared to the real benchmark accuracy of 89.7%. The adverse effect on classifier robustness when reducing the simulation fidelity by altering the motor speed sampling is shown.