Zirui Guo
"UAV-M4: A Multi-Scenario Anti-UAV Multi-Object Tracking BenchmarkAbstract:UAV-M4 is a high-fidelity benchmark dataset designed for vision-based multi-object tracking (MOT) in anti-UAV scenarios. Unlike existing MOT benchmarks that primarily focus on pedestrians or vehicles, UAV-M4 targets homogeneous UAV swarms executing nonlinear tactical maneuvers, a challenging domain where conventional trackers relying on linear Kalman filtering suffer from severe identity switching.The dataset comprises 11,850 frames captured at 30 FPS with a spatial resolution of 640x480 pixels. Four distinct tactical flight scenarios are included: Curved Lift, Sin-Motion, Turning, and Straight. Each scenario features multiple UAVs of the same model flying freely in 3D space, exhibiting diverse motion patterns ranging from linear cruising to highly nonlinear sinusoidal oscillation and continuous turning. All frames are densely annotated with bounding boxes and identity labels, enabling rigorous evaluation of detection, motion prediction, and data association components.UAV-M4 is accompanied by FourierTracker, a frequency-domain motion prediction framework that replaces the traditional Kalman filter with global Fourier reconstruction. On UAV-M4, FourierTracker achieves an average IDF1 of 86.88% and reduces total identity switches by 86% compared to state-of-the-art Kalman-filter-based trackers. The dataset also includes benchmark results for 12 popular trackers (SORT, DeepSORT, ByteTrack, OC-SORT, BoT-SORT, StrongSORT, C-BIoU, DeepOC-SORT, SmileTrack, HybridSORT, DiffMOT, and FourierTracker) across all four scenarios.Potential Applications: (1) Benchmarking MOT algorithms under nonlinear motion, (2) Developing motion prediction modules for anti-drone surveillance, (3) Evaluating kinematic association methods for visually homogeneous targets, (4) Training and validating frequency-domain or learning-based trajectory predictors."