Tracking group targets is one of the major challenges in modern ballistic missile defense system. These targets not only move in an analogue pattern but also are adjacent in space. The projections of group targets on the focal plane array are no longer points but clusters instead, according to the characterization and resolution of the infrared optic sensors. Thus the traditional multi-target tracking methods based on the assumption that each target generates at most one measurement are not fitted any more. In order to realize the group targets tracking, a new filter is proposed. The group targets are treated as a union and the extension of the group is described by an ellipse random hyersurface model. Combined with the Gaussian mixture probability hypothesis density (PHD) filter for extended targets, the group targets are tracked by its centroid states and extensions. With comparisons of the Gaussian inverse Wishart PHD based on the random matrix, the proposed method outperforms the latter one in extension estimation as well as centroid state estimation.