Stacked BiLSTM - CNN for Multiple label UAV sound classification

Abstract

Recently the detection of drones using acoustic data has attracted the interest of researchers, because it is less expensive than other traditional methods. By using acoustic signature we can perform binary classification of UAVs, moreover we can identify if the drone has a load or not. Detection of UAVs with an additional load in the restricted and crowded areas is considered as an effective protection system. This paper considers Multiple label UAV sound classification task using LSTM-CNN architecture. The proposed architecture is composed of Stacked Bidirectional LSTM and CNN, which were learned on representations of the short-term power spectrum of UAV sounds (MFCCs). The results of our experiment show higher accuracy by using a combination of Stacked BiLSTM and CNN rather than using these architectures separately.