Crowd scenes analysis using multiple sliding windows classifiers and Histogram of Oriented Gradient

In recent years many research works have been devoted either to anomaly detection or anomaly classification. However, very few of them address both of them simultaneously. In this paper, we introduced a new method not only to detect and localize the abnormalities in crowded scenes but also to determine the class of abnormality. In This work, we used Histogram of Oriented Gradient to extract the features. Afterwards, we developed a model for each abnormality class based on structured output logistic regression. Using template matching scheme, those regions with maximum detection scores will be chosen as regions which contain abnormality. Aiming to increase model's precision, an iterative hard negative mining has been utilized. Such method was not applicable unless we had general and application free definition for abnormality. Regarding this, we defined a general abnormality definition. The proposed approach shows significant improvements in results over other state-of-the-art approaches.