Bayesian Generative Model Based on Color Histogram of Oriented Phase and Histogram of Oriented Optical Flow for Rare Event Detection in Crowded Scenes

In this paper, we propose a new method for rare event detection in crowded scenes using a combination of Color Histogram of Oriented Phases (CHOP) and Histogram of Oriented Optical Flow (HOOF). We propose to detect and filter spatio-temporal interest points (STIP) based on the visual saliency information of the scene. Once salient STIPs are detected, the motion and appearance information of the surrounding scene are extracted. Finally, the extracted information from normal scenes are modelled by using a Bayesian generative model (Latent Dirichlet Allocation). The rare events are detected by processing the likelihood of the current scene in regard to the obtained model. The proposed method has been tested on the publicly available UMN dataset and compared with different state-of-the-art algorithms. We have shown that our method is very competitive and provides promising results.