Action recognition is an interesting topic in the fields of computer vision and machine learning by the promise of many applications. This topic is still challenging due to some problems, such as environmental change and camera pose variation. In this paper, we present an action recognition method to overcome these problems. Our proposed method introduces a new camera pose-independent motion feature which includes binary ORB (oriented FAST and rotated BRIEF) feature and rotated-normalized optical flow feature extracted at the position of ORB keypoint. The obtained features in each action video are then learned by latent Dirichlet allocation topic model to give action label for that video. We validate the proposed method by conducting experiments in a Unity3D simulation of the operating room.

computer vision, machine learning, action recognition, operating room.