Small Drone Detection from a Moving Camera using On-Device ROI Inference

Abstract

Enabling safe autonomy in unmanned aerial systems (UAS) requires the ability to detect small drones. This is challenging because small drones have a low signal-to-noise ratio (SNR), and limited resources on aerial platforms require lightweight detection models with low-resolution input. We propose an active tracking system that enhances small drone detection by using predictive region of interest (ROI) placement to focus inference resources prior to frame capture. To guide the ROI selection, an extended Kalman filter (EKF) predicts the target's state by incorporating past detections with the camera's own velocity and acceleration, compensating for ego-motion. This approach is evaluated on flight test data captured by a FLIR Firefly DL camera with a MobileNetSSD object detector. Results show an increase in average precision from 20.1\% to 67.0\% when compared to full-frame inference. We validate that predictive ROI tracking with ego-motion compensation enables small object detection on resource-constrained platforms.