Human interaction recognition from distance signature of body centers during time

This paper proposes a simple spatial feature combined with temporal characteristics to classify human interactions from surveillance cameras, which are far from the action scene. For the first stage, data is collected from a horizontal view. Then, the history of distance between two persons is stored during time as a temporal feature called distance signature. We use Spatio-Temporal Interest Points (STIP) to track the body parts and calculate an average kinetic energy for the video sequence. By combining these spatial and temporal features, we use some classification methods to evaluate the features and compare the video classes. Experimental results demonstrate the advantage of the proposed method. Total accuracy of 81%, 84.85% and 93.8% are achieved for DTW, PNN and SVM classifiers, respectively.