Sparse composition of body poses and atomic actions for human activity recognition in RGB-D videos