Cellphone-based sUAS range estimation: a deep-learning classification and regression approach
Anthony C. Brunson, Ryan D. Clendening, Richard Dill, Brett J. Borghetti, Brett Y. Smolenski, Darren M. Haddad, Douglas D. Hodson · 2024
Small Unmanned Aircraft Systems (sUAS) are accessible platforms that pose security threat. These threats warrant affordable and accurate methods for tracking sUAS. We apply a novel approach to estimate the sUAS range using neural networkbased solutions by processing cell phone acoustic recordings, without requiring statistical methods like TDoA or DoA. The data comes from twenty-eight cellphones recording of three different sUAS that fly over the devices. We conduct three experiments as a part of this research. In the first two experiments from [1], the audio data is converted into 0.5s Mel-spectrograms frames and 0.5s raw audio frames, to separate predictions into four range classes. We sequester the data into an 80/20 training test split. The 2DCNN architecture outperforms the other architectures (1DCNN and 2DCRNN). The 2DCNN is then retrained to generalize the sUAS range across various sUAS types to achieve an average Macro-F1 score of 0.7492. In the third experiment, the audio is transformed into 0.1s Mel Frequency Cepstral Coefficient (MFCC) frames to predict the actual distance in meters that the sUAS is from the audio source. A 2DCNN architecture is created that is tested with regression to predict the actual distance in meters that the sUAS is from the audio source. In all scenarios, truth values are calculated from the Euclidean distance between the sUAS and a cell phone. The results show that deep-learning-based sUAS-ranging with cellphones is an effective and low-cost method for accurately tracking sUAS.