A human action recognition scheme based on spatio-temporal variation of region of interest in horizontal and vertical direction

In this paper, human action recognition scheme based on spatio-temporal variation pattern in region of interest (ROI) is proposed. In order to handle the multi-class human action problem in the proposed method two level classification is carried out. In the first level, the motion history image (MHI) is constructed based on optical flow vector, which is capable of classifying static and dynamic human actions because of its distinguishable spatio-temporal variation pattern. In the second stage, further classification of static and dynamic actions are performed. For static action classification, using the pixels inside ROI of MHI, some statistical parameters such as, standard deviation and mean along both vertical and horizontal directions are investigated and found very suitable in distinguishing different classes. However, for dynamic action classification, in view of analysing each frame separately, instead of using MHI, Gaussian mixture model (GMM) based ROI detection algorithm is used which provides detail oriented silhouette of a dynamic object. In order to extract features, first centroid based ROI shifting is performed and then it is segmented into a set of rectangular regions. From the extreme values of the pixels residing inside each region of ROI, within their respective temporal segments, representative features, such as maximum, minimum and mode are computed. It is found that the proposed action recognition scheme not only offers very low computational burden but also can provide satisfactory classification performance even by using simple Euclidean distance based classifier in leave one out cross validation technique.