People Counting and Human Detection in a Challenging Situation

Reliable people counting and human detection is an important problem in visual surveillance. In recent years, the field has seen many advances, but the solutions have restrictions: people must be moving, the background must be simple, and the image resolution must be high. This paper aims to develop an effective method for estimating the number of people and locate each individual in a low resolution image with complicated scenes. The contribution of this paper is threefold. First, postprocessing steps are performed on background subtraction results to estimate the number of people in a complicated scene, which includes people who are moving only slightly. Second, an Expectation Maximization (EM)-based method has been developed to locate individuals in a low resolution scene. In this method, a new cluster model is used to represent each person in the scene. The method does not require a very accurate foreground contour. Third, the number of people is used as a priori for locating individuals based on feature points. Hence, the methods for estimating the number of people and for locating individuals are connected. The developed methods have been validated based on a 4-hour video, with the number of people in the scene ranging from 36 to 222. The best result for estimating the number of people has an average error of 10% over 51 test cases. Based on the estimated number of people, some results of the EM-based method have also been shown.