Pseudo-3D Residual Networks Based Anomaly Detection in Surveillance Videos

We proposed a deep multiple instance learning framework for anomaly detection in surveillance videos. We used training videos with video-level lables and trained the models with multiple instance learning (MIL) methods. We improved the Pseudo-3d ResNet network by adding a batch normalization (BN) operation after each convolution layer, and used the improved network as the feature extractor. In order to achieve better anomaly detection performance, we used the 3-layer full-connected neural network with 50% Dropout regularization as the classifier to obtain anomaly scores, which between 0-1. And the results are compared with detection results of the method with 50% DropConnect regularization. The performance of the proposed approach is evaluated on a large-scale dataset with video-level labels. Experimental results demonstrate that the proposed method in this paper further improves the accuracy of abnormal behavior detection and is more attuned to practical applications as compared to the state-of-the-art approaches.