Applying usability assessment method for surveillance video anomaly detection with multiple distortion

With the extensive deployment of surveillance cameras, video anomaly detection (VAD) is commonly employed to various practical scenarios such as subway stations, parks, and roads. However, the surveillance camera can be easily influenced by weather and hardware degradation during data collection, resulting in information loss. Insufficient information will lead to a decrease in accuracy and credibility for anomaly detection. Accurately measuring the impact of information loss on anomaly detection can be helpful in practical application, and provide reliable application scheme of surveillance data. Therefore, we construct a dataset which contains surveillance data with multiple distortions. Based on the dataset, sufficient reliable data can be provided to measure the impact of data quality for anomaly detection methods. On the basis of the impact of data quality on anomaly detection, thresholds have been designed for data screening to improve the performance of anomaly detection. Finally, an image usability assessment (IUA) method was proposed to accurately screen surveillance data via the designed thresholds. Experimental results demonstrate that the constructed dataset was reasonable and reliable. The proposed IUA method can accurately screen the data to improve the performance of VAD methods, and meet the requirements of practical application scenarios on surveillance data. The dataset has been open-sourced at https://github.com/dart-into/MultipleDistortionDataset.