Action recognition technology has many application fields including gaming, human-computer interaction, security, virtual reality, etc. The technology has recently advanced and simplified by the introduction of real time depth cameras. The most famous depth camera, Kinect, has shown its success in human action recognitions by supervised learning methods. However, those method require a large number of motion capture data as the ground truth of each skeleton points. This paper proposes a human skeleton extraction method without using reference data. The method extracts the minimum number of articulation points instead of full skeleton data. The articulation points are used as an real-time interface for 3D interaction games.