Human Interaction Recognition Using Improved Spatio-Temporal Features

Human Interaction Recognition (HIR) plays a major role in building intelligent video surveillance systems. In this paper, a new interaction recognition mechanism has been proposed to recognize the activity/interaction of the person with improved spatio-temporal feature extraction techniques robust against occlusion. In order to identify the interaction between two persons, tracking is necessary step to track the movement of the person. Next to tracking, local spatio temporal interest points have been detected using corner detector and the motion of the each corner points have been analysed using optical flow. Feature descriptor provides the motion information and the location of the body parts where the motion is exhibited in the blobs. Action has been predicted from the pose information and the temporal information from the optical flow. Hierarchical SVM (H-SVM) has been used to recognize the interaction and Occlusion of blobs gets determined based on the intersection of the region lying in that path. Performance of this system has been tested over different data sets and results seem to be promising.