A hybrid deep network based approach for crowd anomaly detection

In this paper, we present a hybrid deep network based approach for crowd anomaly detection in videos. For improved performance, the proposed approach exploits deep and handcrafted features. The proposed approach extracts spatial and temporal deep features from video frames using two resnet101 models. In order to enhance the deep features discrimination between normal and anomalous events, we perform smoothing of their Euclidean distance values for consecutive frames. For a handcrafted feature that describes the high level motion at the frame level, we compute gradient sum of the frame difference of consecutive video frames. Two deep features and one handcrafted feature of the training frames are used to train three one class support vector machines (OCSVMs). A frame is classified as anomalous performing decision combination of three OCSVMs. Experiments reveal that the proposed approach achieves high accuracy on UMN crowd anomaly dataset. On a more challenging PETS 2009 dataset the proposed approach achieves comparable performance to existing approaches.