Estimating face direction from wideview surveillance camera

In this paper, we propose a facial direction estimating system from low resolution facial images captured by surveillance camera. The proposed system first detects moving objects by background subtraction, then detects pedestrians by using histograms of oriented gradients (HOG) and support vector machine (SVM). After that, it trims head area by template matching and finally estimates facial direction by using another SVM. Experimental results show that, when the SVM learns fluctuated facial images, it achieve more than 96% estimation accuracy for facial database images. Experimental results of actual surveillance camera images show that, by learning fluctuated images, estimation accuracy of facial directions improves from 34% to 38%. When we tolerate estimation error within 30 degrees, it can achieve 72% estimation rate.