Crowd density estimation based on image potential energy model
Guogang Xiong, Xinyu Wu, Jun Sheng Cheng, Yen-Lun Chen, Yongsheng Ou, Ying Liu · 2011
Reliable estimation of crowd density in public plays an important role on intelligent surveillance in recent years. There have been a lot of research on people counting; however, most of them only consider crowd with slight occlusions and their algorithms usually accompany with high computational complexity. In this paper, we present a simple model based on image potential energy to estimate the crowd density. The image potential energy is inspired by gravitational potential energy. Based on the facts that the pixels related to the object on the image plane are fewer if the object is farther away from the camera and the farther objects appear closer to the origin of the image plane, we define the image potential energy on the image plane. The main characteristics of the model is that the image potential energy related to objects is almost invariable no matter how far away the object being from the camera. The potential energy model can deal with severe occlusions with low computational complexity. It is adaptive to low and high density of crowd in public scenes. When the crowd density is below 10, the model accuracy rate is about 80% and the error is about 1 people count for a series of frames. When the crowd density varies from 10 to 40, the crowd density changes very fast, we can't make accuracy analysis as in low crowd density; however, for one single frame, the error rate is below 7% while the average error varies from 1 to 3 in the experiments.