Abstract: Aiming at the moving object tracking problem of satellite video, a robust method of feature description and matching based pointtarget tracking is proposed. The idea of correlation filter is introduced. Firstly, Hu invariant moments and median filtering of the sample set are used to establish the tracking template of the target and describe the target features. Then, the dimension of the decision area of target tracking is reduced and the Hu confidence model of the judgment area is established. Finally, the target tracking is achieved by finding the maximum value of the tracking confidence map by using the fast FFT derivation correlation method. Moreover, the robustness of the algorithm is improved by using Kalman filtering to assist and optimize the tracking process. The experimental video data are taken from SkySat1 and Jinlin1 satellites. The tracking experiments of 5 point targets show that the proposed method achieves good results in the satellite video, and that the accuracy is better than 90%. The target is not lost in the whole tracking process and the tracking trajectory is smooth. In addition, for the 13×13 judgment area, the processing speed of the confidence map is about 5 times higher than the conventional correlation method. This method can provide a technical basis for achieving fast and reliable tracking of point targets in satellite video.

Key words: Hu correlation filter, satellite video, pointtarget tracking, fast correlation method, Hu confidence model, dynamic information, trajectory extraction