Real-World Anomaly Detection Using Deep Learning

In this paper, we have carried out a comparative study on two deep learning models for detecting real-world anomalies in surveillance videos. Anomalous event is the one which deviates from the normal behavior. The anomalies considered are related to thefts such as robbery, burglary, stealing, and shoplifting. A framework is set up using supervised learning approach to train the models using theweakly labeled videos. The deep learning models considered areVGG-16 and inceptionmodel which are trained with both anomalous and normal videos to detect any anomalous activity in the video frame. UCF-Crime dataset is used which comprises long, untrimmed surveillance videos. The deep learning models are evaluated using various metrics. The experimental results show that the Inception V3 model performs significantly better in detecting the anomalies as compared to theVGG-16 model with an accuracy of 94.54%.