Evaluating a Motion-Based Region Proposal Approach with Background Subtraction Methods for Small Drone Detection

The detection of drones in complex and dynamic environments poses significant challenges due to their small size and background clutter. This study aims to address these challenges by developing a motion-based pipeline that integrates background subtraction and deep learning-based classification to detect drones in video sequences. Two background subtraction methods, Mixture of Gaussians 2 (MOG2) and Visual Background Extractor (ViBe), are assessed to isolate potential drone regions in highly complex and dynamic backgrounds. These regions are then classified using the ResNet18 architecture. The Drone-vs-Bird dataset is utilized to test the algorithm, focusing on distinguishing drones from other dynamic objects such as birds, trees, and clouds. By leveraging motion-based information, the method enhances the drone detection process by reducing computational demands. Results show that ViBe achieves a recall of 0.956 and a precision of 0.078, while MOG2 achieves a recall of 0.857 and a precision of 0.034, highlighting the comparative advantages of ViBe in detecting small drones in challenging scenarios. These findings demonstrate the robustness of the proposed pipeline and its potential contribution to enhancing surveillance and security measures.