Real-Time Deep Anomaly Detection: An Overview of Benchmark Datasets and Performance Metrics

Abstract

Safeguarding lives and properties in public places is one of the key components of the smart city. Therefore, the Intelligent Video Surveillance System (IVSS) can use video anomaly detectors to detect various anomalous activities using live streaming of video. Anomaly/Abnormal Activities are those actions that occur at unusual locations/periods. Activities such as fighting, vandalism, riots, theft, wrong U-turns, and road accidents are examples of abnormal activities. Abnormal activities pose potential danger to the well-being of people in a smart city, which necessitates prompt detection. Various deep-learning algorithms are used to detect anomalies in videos. To evaluate the quality of the generated results of these algorithms, appropriate datasets, evaluation metrics, and hyperparameter optimization are needed preferably combined in one research work. To the best of our ability, there is no research survey that purposely delve on these three (3) mentioned concepts in one piece of writing. The research centres on examining prominent datasets and evaluation metrics used in Video Anomaly detection(VAD). It gives an overview of the state-of-the-art (SOTA) Datasets and evaluation metrics used in assessing the performance of VAD methods, as well as hyperparameter tuning which provides the best result on the dataset in a realistic time frame (time-to-accuracy). Our findings reveal that the UCF Crime is the most suitable dataset for VAD as it meets all set criteria and the area under curve(AUC)/receiver operating characteristics(ROC) is often used for evaluating the performance of models in VAD. Finally, issues and prospects were given on the topic. A fully implemented IVSS will go a long way in providing safety in public places and transport systems through prompt notification of anomalies to prevent loss of lives and properties.