Anomaly detection method of surveillance video based on global-local information

The problem of detecting anomalies in videos is usually regarded as a Multi-Instance Learning problem under the weak supervisory approach. While existing methods have shown superior performance, most of them ignore the connection between global and local information in the video. They only model normal or abnormal segments with the highest anomaly score. As a result, the ability to generalize such hard-to-distinguish abnormal behaviors as small-amplitude abnormal events or large-amplitude normal events (e.g., accelerated running or stealing) is poor. To address this problem, a global-local based hard-to-categorize anomaly detection network is proposed. This network utilizes channel features, spatio-temporal features and motion information features of video clips to capture global and local feature information in videos to distinguish abnormal and normal events. Secondly, a global-local based anomaly ranking loss function is designed using a Multi-Task Learning approach, which increases the focus on hard-to-distinguish samples and reduces the over-learning of easy-to-categorize samples. In addition, based on the loss function, a score range scaling strategy is proposed, by which the classification boundary between anomalous and normal examples is shortened, making the model more sensitive to less obvious anomalous samples. It shows higher robustness on hard-to-categorize anomalous scenarios such as light interference, small motion changes, etc. On the UCF-Crime dataset, the AUC reaches 83.9 %, an increase of 1.6 % over the original model. It provides new ideas in the field of anomaly detection.