A Novel Attention Heat Map-Based Video Anomaly Detection Framework

In video surveillance systems for the purpose of public safety, automatic anomaly detection is a vital responsibility. The goal of anomalous activity recognition is to identify the frames with various spatial information. In the proposed methodology we presented a framework where we can classify and localize rare events by using trainable features only through the global average pooling layer rather than flattening the model, using the pre-trained model. This model is trained only on the global average pooling layer while the other layers are non-trainable. Extracted features are further connected to dense layers to make predictions as anomalies or normal images. We first determined all images classified as anomalies by the model to localize anomalies and save those images. Using a heat map with peak value and threshold we locate different anomalies in the image using weights of the last two layers of the trained model. Our method adds the advantage of both frame-level and pixel-level performance where we classify and locate the anomalies. We evaluated our network performance on two standard datasets that outperform the state of art methods by achieving an accuracy of 93.7% for UCSD ped1 and 95.4% for ped2 datasets respectively.