HUMAN ACTION RECOGNITION

In this project, we develop an algorithm that recognize different human actions from videos. We tackle the problem of human silhouette extraction and present three different methodologies or techniques that adress the human detection from an image. Our particular project first trains a system utilizing the Optical Flow for tracking the actions. With the Flow Farneback method we track for some sequences the pixels from the regions inside a bounding box which contains the actor. This process results in thousands of tracks addressed as visual words. The visual words are lately clustered using the k-mean method to gain a minimal set of visual words describing actions compactly. For our particular, we generate a set of visual words for the videos to probe. Both systems are compared with each other showing in the classification experiments the importance of finding an appropiate approach between both, small and large motions.