Lightweight machine learning models for drone detection using acoustic and optical features

Abstract The widespread deployment of drones has triggered major concerns over privacy and security, creating a demand for robust anti-drone systems (ADS). A critical component of ADS is the detection unit, which identifies drones in an unauthorized area. Recently, various statistical and machine/deep learning methods have been developed for drone detection units. Statistical methods are traditionally applied which often suffer from uncertain thresholds under varying noise distributions. While deep learning-based methods are highly popular, they frequently face challenges related to high computational complexity. This study explores the potential of low-complexity machine learning (LCML) models, including logistic regression model (LRM), support vector machines (SVM), and random forest algorithm (RFA) for drone detection using acoustic and optical features. For drone detection, hybrid features such as histogram of oriented gradients and ResNet features are extracted from images are used. HOG features are derived from log-mel spectrograms of drone acoustic signals. The LCML models are assessed using various performance metrics for binary classification, with LRM demonstrating the best results that achieves 97% accuracy with optical features and 98% accuracy with acoustic features. In addition, LRM exhibits the lowest training complexity of $$O(N \cdot d \cdot i)$$ and inference complexity of O ( d ), making it suitable where limited computational resources are available. These findings suggest that LRM is the favourable LCML model for real-time detection unit for an ADS that offers a balance between accuracy and inference complexity.