People counting using iterative mean-shift fitting with symmetry measure
Li Chen, Tao Ji, Yap‐Peng Tan, Kap-Luk Chan · 2007
This paper presents a new method for counting the number of persons from video images. Conventional people counting methods can be classified into the learning-tracking based techniques. They either require elaborated human model learned using AdaBoost or sophisticated tracking algorithms by particle filtering. The proposed algorithm performs people counting by segmenting group of people in a cluttered scene into individuals. By assuming human body is bilaterally symmetric, the proposed method first determines a probability map of symmetry using a local energy function. Taking the highest symmetry pixel as the initial position, our method then employs the mean-shift technique to fit each person in the foreground. The procedure of mean-shift fitting is repeated until all the foreground regions are exhausted. The advantages of the proposed method lie in its simplicity and capability of counting multiple human targets without stringent learning and tracking constraints. Experimental results show that the proposed method is robust and efficient in handling real-life video data of different scenarios.