Human Action Recognition by Employing DWT and Texture

Human action recognition is a very challenging task due to the great variability with which different people may perform the same action. It involves in the development of applications such as automatic monitoring, surveillance, and intelligent human–computer interfaces. We propose an action recognition scheme to classify human actions based on positive portion using template-based approach from a video. We first define the accumulated motion image (AMI) using frame differences to represent the spatiotemporal features of occurring actions. Then, the direction of motion is found out by computing motion history image (MHI). Texture and spatial information are extracted from AMI and MHI using (LBP) local binary pattern and (DWT) discrete wavelet transform, respectively. The detection of object and extraction of moving objects are done by feature extraction over LBP and DWT. The feature vectors are computed by employing the seven Hu moments. The system is trained using nearest neighbor classifier, and the actions are classified and labeled accordingly. The experiments are conducted on Weizmann dataset.