Tiny Drone Detection from Videos with Tracking-Based Temporal Lost Compensation

This paper proposes tiny drone detection from videos with tracking-based temporal lost compensation. Our method first detects objects from each frame in a video using neural networks based on Faster R-CNN with a Gaussian receptive field based label assignment (RFLA). It then tracks the detected objects between consecutive frames by fast and robust generic multiple-object tracking (FRoG-MOT). To further reduce miss detection, our method compensates for lost targets by the appearance-based association of tracklet predictions. We evaluate our method using the training dataset provided by the Drone vs. Bird Detection Challenge. Experimental results show that our method improves average recall by 1.7 points compared with the baseline object detection, keeping comparable precision. With improved detection performance for tiny drones, we detect distant drones continuously without temporal miss detections, ensuring the safety of essential facilities. This method is the main algorithm in our submission to the 8th Drone vs. Bird Challenge in IJCNN 2025.