A novel robust tracking algorithm for anti-UAV based on dynamic similarity scale estimation and adaptive occlusion-aware

Unmanned aerial vehicle (UAV) detection and tracking methods are imperative to preventing UAV from threatening public safety. However, the various challenges, such as dynamic random scale variation, occlusion, and the reappearance of targets, are usually encountered during tracking UAV. Therefore, a long-term Anti-UAV tracking algorithm, which incorporates a kernelized correlation filter (KCF) and YOLOv7 re-detection module, is proposed to deal with above-mentioned problems. A scale estimation module, which utilizes the binary image similarity metric, is firstly proposed to estimate the dynamic random scale variation of UAV. Compared with the traditional KCF algorithm, an adaptive occlusion-aware mechanism, which combines the nonlinear smooth shift function and the peak-to-sidelobe ratio, is innovatively designed to handle the problem of the long-term occlusion. To attenuate the influence of the time-variant target and background on the update of the appearance model, the target template update strategy is optimized by adjusting the learning rate adaptively. Moreover, YOLOv7 re-detection module is introduced to address the challenge of UAV reappearance during long-term tracking. Experiments are conducted on a self-built UAV dataset, and the results demonstrate that the success rate and precision of the proposed method are respectively increased by 15.3% and 18.9% compared with the baseline KCF algorithm. Furthermore, the proposed method could operate at a speed of 132 FPS.