Acoustic localization and identification of drones with a disturbance source
Résumé
Due to their low cost and their easy use, drones are increasingly used as tools to threaten valuable assets in both military and civil domains. Thus systems have been developed to localize and/or identify them before their neutralization. They can rely on radio frequency, radar signal or even video recording. However these modalities suffer from limitations like near-field detection or weather conditions. Besides, drones do produce acoustic noise during their flight due to their propellers. Consequently, the proposed work uses acoustic waves emitted by drones, array processing and machine learning to localize them and identify them as drones. As multiple acoustic sources can be present within the nearby environment, this work focuses on localization and identification of a drone in the presence of a disturbing acoustic source. An experiment has been performed with a drone flying above a 81-microphones acoustic array of 1 m diameter while a motionless loudspeaker emits noise. Signals recorded on the array are cut into successive blocks of 200 ms to allow an ?instantaneous? identification and tracking of the different sources. MUSIC algorithm is used on each block to estimate the angular directions of arrival of the sources, using local maxima of the localization map. Then the focused signals are built for each source and classified as drone or not thanks to a machine learning model (SVM) learned previously on similar data. Results show that the classification task depends on the signal to noise ratio between the two sources and on their relative position. A final drone trajectory can be estimated thanks to the classification step that allows to tag every detected drone within each block along time.
Origine : Fichiers éditeurs autorisés sur une archive ouverte