Pedestrian analysis and counting system with videos

Abstract

Reliable estimation of number of pedestrians has played an important role in the management of public places. However, how to accurately count pedestrians with abnormal behavior noises is one challenge in such surveillance systems. To deal with this problem, we propose a new and efficient framework for pedestrian analysis and counting, which consists of two main steps. Firstly, a rule induction classifier with optical-flow feature is designed to recognize the abnormal behaviors. Then, a linear regression model is used to learn the relationship between the number of pixels and the number of pedestrians. Consequently, our system can count pedestrians precisely in general scenes without the influence of abnormal behaviors. Experimental results on the videos of different scenes show that our system has achieved an accuracy of 98.59% and 96.04% for the abnormal behavior recognition and pedestrian counting respectively. Furthermore, it is robust against the variation of lighting and noise. © 2012 Springer-Verlag.