Método espacio-temporal para el reconocimiento de acciones humanas en el espacio canónico

The recognition of human actions is a very active research field in computer vision, where efforts are presently focused on the detection of human behavior in real-time video. In this paper, we present a novel spatio-temporal algorithm, called the Motion Vector Flow Instance (MVFI), for classification of actions in videos. We show the results of applying this algorithm to two public datasets, "KTH" and "MILE" that contain scenes of human actions with different recording conditions (multiple camera angles, lighting, different clothes, and video quality). The MVFI spatio-temporal template encodes information about the speed and direction of human motion from the optical flow vectors obtained within each video frame. Then, by using supervised learning, MVFI images are projected into a canonical vector space and decision boundaries are determined for various actions by using a support vector machines (SVM) algorithm. Thus, in this paper, we demonstrate that our method is robust for detecting human actions across different datasets and provides real-time recognition.