Hybrid Active Contour Shape Features for Human Action Recognition using Deep Learning Model

Abstract

Over the last decade, Human Action Recognition (HAR) has to turn out to be an important research domain for several applications. In fact, “HAR is the problem of classifying an action and assigning it a label of action class”. When an action is performed by a human, her / his body parts show different movements of the body. For detecting these movements and the performed actions, the researchers have to model a video system. The majority of the current works carried out on HAR are focused on videos; however, they include diverse constraints, which make the action recognition process a complex task. Hence, this paper intends to design a novel approach for HAR from the video source. Initially, the keyframes are extracted from the input video and movement density is computed for the keyframes. Subsequently, the region-based active contour-based segmentation is carried out and region based intensity map is generated. Moreover, the feature extraction plays a major role in this work, which is carried out on a hybrid based scenario. Specifically, the active contour generates the edge point features and on the other side, the midpoint features get extracted from the hierarchical centroid model. Henceforth, the features extracted (combined midpoints and edge points) are subjected to Deep belief Network (DBN) for HAR. Finally, the performance of implemented work is compared and proved over the conventional models with respect to error analysis.