Counting Pedestrian in Crowded Subway Scene

When the high occlusion occurs in crowded scene, face detection is a better substitute for detecting pedestrian. In this paper, we present a novel crowd analysis method based on discriminative descriptor of faces and support vector machine (SVM) ensemble. Through manipulating the input features in the same sample set, the different input features of faces are extracted to train two SVM classifiers. The classification scores of two generated classifiers are combined adaptively to make a collective decision. The first SVM, as the principal classifier gives out most of face hypotheses, while the second SVM serves as secondary one to rejecting the false positive. We present experiment to test the proposed method in crowded subway video, and the result shows that the SVM ensemble outperforms the single SVM in counting the pedestrian.