A Benchmark and Simulator for UAV Tracking Matthias Mueller, Neil Smith, and Bernard Ghanem King Abdullah University of Science and Technology (KAUST) Acknowledgements. Research reported was supported by competitive funding from King Abdullah University of Science and Technology. UAV V Tracking g Benchmark Dataset ¾ 123 aerial sequences, >110,000 frames ¾ Seven times larger than VIVID ¾ Second largest object tracking dataset Evaluation ¾ Attributes (aerial tracking) ¾ Spatial robustness, ¾ Sensitivity to frame rate ¾ Long-term tracking Results on OTB100, UAV123 and UAV20L 64 35 11 TC128 51 38 21 UAV123 72 21 7 OTB100 easy medium hard Difficulty is calculated as mean of precision and success score of the best tracker per sequence. Easy >90%, Medium >50%, Hard <50%. 64% 16% 9% 39% 49% 89% 55% 39% 23% 59% SV ARC LR FM OCC Attribute Comparison OTB100 UAV123 109 68 48 28 33 73 30 21 31 60 70 39 SV ARC LR FM FOC POC OV BC IV VC CM SOB Attribute Distribution Simulator r (Unreal al Engine e 4 4 4) Highlights ¾ UAV Physics Simulation ¾ Visual servoing system ¾ Frame capture and flight logging ¾ MATLAB/C++ integration of trackers Synthetic Sequence Generation ¾ Custom depth maps for any mesh/object ¾ Automatic ground truth annotation Live Tracking with Feedback ¾ Planned path or manual control of target ¾ UAV is controlled by tracking algorithm ¾ Live visual feedback and novel evaluation Qualitative Visualization ¾ Generate UAV trajectories from log files ¾ User-defined camera views ¾ VR integration with HTC Vive Benchmark and Simulator available at: https://goo.gl/LBC4zU