Salient Attention Model and Classes Imbalance Remission for Video Anomaly Analysis with Weak Label

摘要

Recently, weakly supervised anomaly detection has got more and more attention. In several security fields, realizing what kind of anomaly happened may be beneficial for security person who have preparation to deal with. However, lots of studies use global features aggregation or topK mean, and it exists feature dilution for anomaly. An attention model is proposed to generate the segment scores, i.e. we propose a salient selection way based on attention model to efficiently detect and classify the anomaly event. With these selected highlighted features, graphs are constructed. Graph convolutional network (GCN) is powerful to learn the embedding features, anomaly event can be expressed more strongly to classify with GCN. Because normal events are common and easy to collect, there is a problem that the normal and abnormal data are imbalance. An abnormal-focal loss is adapted to reduce influence of large normal data, and augment the margin of normal and different anomaly events. The experiments on UCF-Crime show that proposed methods can achieve the best performance. The AUC score is 81.54%, and 0.46% higher than state-of-the-art method. We obtain 58.26% accuracy for classification, and the normal and anomalies are separated better.