Abnormal Behavior Detection in Crowded Scenes Using Density Heatmaps and Optical Flow

Crowd behavior analysis is an arduous task due to scale, light and crowd density variations. This paper aims to develop a new method that can precisely detect and classify abnormal behavior in dense crowds. A two-stream network is proposed that uses crowd density heat-maps and optical flow information to classify abnormal events. Work on this network has highlighted the lack of large scale relevant datasets due to the fact that dealing and annotating such kind of data is a highly time consuming and demanding task. Therefore, a new synthetic dataset has been created using the Grand Theft Auto V engine which offers highly detailed simulated crowd abnormal behaviors.