1: \begin{abstract}

2: In this paper, we propose a novel method for representation and classification of two-person interactions from 3D skeleton sequences. 

3: The key idea of our approach is to use Gaussian distributions to capture statistics on $\mathbb{R}^n$ and those on the space of symmetric positive definite (SPD) matrices.  

4: The main challenge is how to parametrize those distributions.   

5: Towards this end, we develop methods for embedding Gaussian distributions in matrix groups based on the theory of Lie groups and Riemannian symmetric spaces.   

6: Our method relies on the Riemannian geometry of the underlying manifolds and has the advantage of encoding high-order statistics from

7: 3D joint positions.  

8: We show that the proposed method achieves competitive results in two-person interaction recognition on three benchmarks for 3D human activity understanding.   

9: \end{abstract}

10: