Evaluation of acoustic detection of UAVs using machine learning methods

In the past years, small unmanned aerial vehicles have increasingly become a hard to defend against threat to both military and civilian infrastructure. Both DIY and COTS UAVs are difficult to detect in a realistic environment, particularly in a cluttered and noisy one such as a port facility. In this context the use of active detection systems such as radar, and lidar is limited since it should not adversely interfere with the normal operation of the port, in particular the existing harbour sensors. This leads our interest towards passive multi-sensor detection systems such as electro-optic (EO) and acoustic monitoring. This work investigates the capability of passive acoustic systems to detect small commercial UAVs within the context of a harbour. We use a machine learning approach to detection using real-world data. We collected audio signatures of several different types of commercial off-the-shelf UAVs both in a quiet environment and in a variety of complex real environment. For this we used a directional 4-microphone array composed of readily available audio components. This setup limited our experiment to the audible spectrum, in which motor and propeller noise are the main characteristics used to distinguish the UAV from the background sounds. We studied machine learning algorithms typically applied to this category of problems, and implemented a Gaussian Mixture Model (GMM) classifier using the Mel-Frequency Cepstrum Coefficient (MFCC) as a feature representation of the audio data, and apply this to the data collected during our measurement campaigns.