Human action recognition using spatio-temporal features

Abstract

Recognition of human action from video database is a trivial task as they are prone to noise and low resolution. To reduce the disadvantages and to avoid the foreground segmentation and body parts tracking, a new low-level visual feature called Spatio-temporal context distribution feature of interest points is used to describe human actions. Each action video is expressed as a set of relative XYT coordinates between pair wise interest points in a local region. A locally described word context descriptor(LWWC) is used in which the Spatio-temporal interest points are initially described by Histogram of optical flow (HOF) and Histogram of oriented gradients (HOG) which improves the discriminability of the traditionally used local spatial-temporal descriptors. In the action units of the context-aware descriptors, graph regularized nonnegative matrix factorization is applied which leads to a part-based representation of the video and encodes the geometrical information from them. For the classification kernel SVM is proposed as the kernel weights for each action videos are obtained for each specific action videos. These kernel weights are trained along with the action units in the SVM classifier to increase the accuracy of classification.