Abnormal crowd event detection by fusing saliency information and social force model

Abnormal event detection plays an important role in intelligent video surveillance. A new abnormal behavior detection algorithm is presented by fusing spatiotemporal features. We first extract SI as the feature representation in the spatial domain. Then, by combining the high precision optical flow algorithm with social force model, we calculate the interaction force as the feature representation in the temporal domain. A novel motion feature descriptor, i.e., Histogram of Interaction Force (HOIF) is proposed, which is combined with SI as temporal-spatial features to be input to the Support Vector Machine (SVM) to identify the crowd events. The effectiveness of the proposed algorithm is put to test on the UMN dataset, and the experimental results indicate that the presented method offers more reliable performance than some existing algorithms in terms of accuracy and robustness.