AN MLP CLASSIFICATION METHOD FOR SOUND-BASED DRONE DETECTION SYSTEM

With the increase in various types and production of drones for multiple purposes, there may be a danger of using drones illegally. This condition harms protected public areas, such as tourist attractions and places of worship, offices, etc. Therefore, it is crucial to detect events or conditions considered detrimental so that security operators can obtain this information and identify the presence of drones. This paper introduces a drone detection method based on sound data using a combination of Log-mel spectrogram and Mel-Frequency Cepstrum Coefficients (MFCC) features. Therefore, this paper’s aim is to offer a deep learning methodology applied to tasks related to drone detection using voice data. Based on experimental results, this MLP-based model detected the presence of drones 549 times and caught non-drones 241 times from the data tested (20% of the dataset of 4,093 data). The accuracy percentage achieved is 96.46%; in the test data, the precision value is 96%. The highest recall percentage and f1-score were achieved in the Drone class (recall = 98%, f1-score = 97%). The confusion matrix from the model evaluation shows that the model made nine times the error of detecting objects as Non-Drone (FN = 1.1%). ICIC International