Deep neural networks algorithms for acoustic drone detection

In 2010, the first commercial drone was presented at Consumer Electronics Show (CES), and since this date drones are becoming increasingly popular in various industrial, commercial, and public-safety areas. However, drones can be used in several illegal activities, and they pose serious challenges especially in highly security-sensitive areas such as airports and nuclear plants. As a consequence, effective counter measures are highly required in order to detect and report a drone flying over such restricted areas. Recent advances and fast developments in the design and implementation of deep learning models lead us to apply them in several recognition tasks such as speech, music, environmental sounds, and image recognition. Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) are two widely used deep learning models, where RNNs show remarkable performance in several sequence data related tasks such as natural language processing applications. On the other hand, convolutional models considerably success in image classification and object recognition for computer vision tasks. The main goal of the thesis is to develop a powerful and efficient deep learning model for acoustic drone detection. In this context, we investigate the results of the drone's sound recognition scheme based on RNNs and CNNs that are trained using our collected dataset; we also investigate the influence of acoustic features extraction techniques such as MFCCs and Mel-Scale Filter Banks on model's classification performance with different sampling rates. It is verified with various experiments that the CNN model with Mel-scale filter banks as a feature extraction technique with 32 KHz unified sampling rate gives the best classification performance. Keywords: Deep Learning, Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Acoustic Features, Mel Frequency Cepstral Coefficients MFCCs, Mel-Scale Filter Banks