Hidden Markov model based drone sound recognition using MFCC technique in practical noisy environments

The development of drones has captured the attention of hobbyists and investors alike; drones now have greater commercial and military applications owing to their relatively small size and ability to fly without an on-board pilot. However, certain drone applications may pose serious threats to public safety. The most important problem to be addressed is the recognition of drones in security-sensitive areas. This paper presents an approach to recognize drones through sounds emitted by their propellers using Mel frequency cepstral coefficients (MFCCs) technique for feature extraction and the hidden Markov model (HMM) approach for classification. In the feature extraction stage, two schemes for feature vectors (one using twenty-four MFCCs and the other using the proposed thirty-six MFCCs) are applied, where additional dynamic information of the features is added in the latter. The classifier based on HMMs is then trained using the extracted features according to different training datasets in order to validate the effect of the number of sound types in each cluster on the recognition rate performance. We perform experiments for drone sound recognition utilizing various training datasets for the purpose of classifier optimization, as well as for the two MFCC schemes that are applied in each trial, using the same training datasets for a fair comparison. The experimental results finally validate the feasibility and effectiveness of our proposed methods with relatively high recognition rates, even in noisy environments.