Drone Sound Classification using Chromagram Features: A Novel Approach

Unmanned Aerial Vehicles (UAVs), or drones, are increasingly equipped with technologies like GPS, LIDAR, radar, and visual sensors, enabling diverse applications such as cinematography, agriculture, surveillance, and leisure. These advanced UAVs are promising for tasks like infrastructure monitoring, emergency aid delivery, search and rescue, data collection in hazardous environments, and military operations. However, their misuse for activities like privacy breaches, smuggling, and security threats raises public safety concerns. Detecting and classifying UAVs is challenging, with current methods using radar, video, acoustic, and radio frequency detection. Machine learning (ML) techniques, particularly Support Vector Machines (SVM), are effective for UAV detection and classification. SVM performance depends on input features, with previous research using features like LPCCs, GTCCs, and MFCCs. Our study seeks to investigate the impact of alternative feature sets, specifically exploring a range of chromagram features along with rolling mean, rolling standard deviation, and exponential weighted moving average to differentiate UAVs from background noise. The integration of these features with a streamlined SVM model serves to concurrently elevate performance and efficiency while maintaining cost-effectiveness. This study also proposes alternative approach like to transform audio into visual inputs for CNNs and employing GNNs with chromagram features as the input. Findings of this study indicate that ML-based strategies outperform the other alternatives. The results demonstrate that strategies based on ML with chromagram features surpass alternative approaches. When considering the array of ML models, such as SVM, Random Forest (RF), k-Nearest Neighbor (KNN), which are trained using a broad spectrum of chroma features in addition to statistical ones, the SVM model exhibits superior performance metrics on public datasets. It eclipses existing top methods for sound classification. Furthermore, our results align with earlier research conducted by Salman et al. (2021).