Video Anomaly Detection Using Encoder-Decoder Networks with Video Vision Transformer and Channel Attention Blocks

Abstract

A surveillance camera has been introduced in various locations for public safety. However, security personnel who have to keep observing surveillance camera movies with few abnormal events would be boring. The purpose of this study is to develop a computerized anomaly detection method for the surveillance camera movies. Our database consisted of three public datasets for anomaly detection: UCSD Pedestrian 1, 2, and CUHK Avenue datasets. In the proposed network, channel attention blocks were introduced to TransAnomaly which is one of the anomaly detections to focus important channel information. The areas under the receiver operating characteristic curves (AUCs) with the proposed network were 0.827 for UCSD Pedestrian 1, 0.964 for UCSD Pedestrian 2, and 0.854 for CUHK Avenue, respectively. The AUCs for the proposed network were greater than those for a conventional TransAnomaly without channel attention blocks (0.767, 0.934, and 0.839).