Human Action Recognition using Optical Flow Estimation

Human Action Recognition is a developing field in computer vision and machine learning. Our aim is to perceive the game being played in the video. There are different procedures like extracting feature vector or keypoints for Human Action Recognition. Optical flow is calculated, which gives the estimation of movement, of a pixel in a sequence of frames. The data set that we are utilizing is UCF Sports dataset. This is a popular human action recognition dataset. Labels can be provided but such labels are for humans and not for machines. The issue is that, if an unlabelled video is given, it is difficult for the machine to perceive the class of the video. In this paper, velocity profile of every pixel in several frames is calculated including magnitude and orientation. Further, orientations are divided into several bins on the basis of explicit results. These bins are used for detailed analysis and conditioning for particular human action depicted in the sports