Abstract
Crowd analysis has numerous applications in crowd safety and security. In order to automate the process of crowd analysis, crowd segmentation is the pre-processing step. In this paper, we propose crowd segmentation framework that extract crowd regions from the background. We can extract crowd regions by employing background modeling and motion segmentation techniques. Since these techniques use motion cues, therefore accumulate false positives in the scenes where the crowd is stationary. In order to avoid using motion cues, we propose a fast and robust crowd segmentation framework that exploits appearance and structure cues to distinguish between the crowd region and background. We train appearance and structure based models separately and then jointly optimized the pre-trained models. To evaluate the performance of our proposed framework, we collect a data set that includes images from different complex scenes. From the experiment results, we observe that our proposed framework achieve superior performance compared to other state-of-the-art methods.