A real-time global-search method for infrared UAV tracking

Tracking unmanned aerial vehicles (UAVs) using infrared cameras is an important technology for the anti-UAV task. Most of the existing trackers adopt the local-search method that crops out an image patch for tracking targets, which is not effective in dealing with challenging situations such as target disappearance and tracking failure. And attaching a global detection module to the local tracker or directly using a global tracker will require too much computation and cannot meet the real time requirement. Therefore, we propose a real-time global-search method for tracking UAVs in infrared videos, dubbed RGT, which is based on a one-stage anchor-free deep framework and introduces an attention module to enhance the ability to discriminate between the target and the background. In addition, according to the multiscale output characteristics in our model, we propose an adaptive multi-scale template updating (AMTU) and a multiscale spatial constraint (MSC) to address the problems of target scale variation and background clutter, respectively. We use the 1st, 2nd and 3rd Anti-UAV datasets for training and testing, and implement comparison experiments with 10 deep trackers. Our algorithm shows excellent performance and can run in real time at 30FPS on an NVIDIA RTX 2080Ti GPU.