Tiny drone-tracking framework using multiple trackers and a Kalman-based predictor

Unmanned aerial vehicles are among the major developing technologies that have many beneficial applications. However, they can also pose a significant threat. Thus, a high demand exists for the development of a surveillance system for drones. However, tracking tiny drones with a moving camera is challenging. First, drones move fast and are usually tiny. Second, images captured by a moving camera have illumination and global motion changes. Moreover, a tracker should perform in real time. To handle these conditions, this paper proposes a tracking framework comprising two trackers, a predictor, and a refinement process. Trackers use motion flow and a histogram feature, respectively. Each tracker locates the region of interest (ROI) of a target. A Kalman filter is also used to estimate the trajectory of the target as the predictor. To locate the target from the two ROIs, the ROIs are compared using the trajectory, and the target is localized in the refinement process. For experiments, a dataset in which a drone flying is captured by a moving camera is used. The experimental results reveal that the proposed method outperforms the existing methods, such as MedianFlow, MIL, Boosting, KCF, and STAPLE regarding the average center error and average success rate.