Pedestrian counting in video sequences based on optical flow clustering

Abstract

The demand for automatic counting of pedestrians at event sites, buildings, or streets has been increased. Existing systems for counting pedestrians in video sequences have a problem that counting accuracy degrades when many pedestrians coexist and occlusion occurs frequently. In this paper, we introduce a method of clustering optical flows extracted from pedestrians in video frames to improve the counting accuracy. The proposed method counts the number of pedestrians by using pre-learned statistics, based on the strong correlation between the number of optical flow clusters and the actual number of pedestrians. We evaluate the accuracy of the proposed method using several video sequences, focusing in particular on the effect of parameters for optical flow clustering. We find that the proposed method improves the counting accuracy by up to 25% as compared with a non-clustering method. We also report that using a clustering threshold of angles less than 1° is effective for enhancing counting accuracy.