STVAD: A Lightweight Spatio–Temporal Attention Network for Video Anomaly Detection

Detecting anomalies within video content continues to pose an alarming challenge in social systems, attributable to the inherently ambiguous nature of abnormal activities and the complex nature of video data. The presence of weakly labeled data and the substantial size of models hinder the widespread application of existing methods in real-world situations, particularly on edge devices with limited resources. This article introduces a lightweight anomaly detection system for videos with a novel spatio–temporal attention module in a weakly-supervised environment. With binary-level annotation of videos, the system’s task is to identify abnormal behaviour from video frames. Existing works focus more on strange objects in the videos. The infrequent occurrence of abnormal events and diversity in daily activities poses challenges in such systems. Exploring the frames in a video for the salient and discriminative features is crucial in focusing and identifying critical areas in a frame. This article explores the spatial and temporal characteristics of a video for the detection of abnormal behaviour. More specifically, we put forward a spatio-temporal attention mechanism to extricate the relevant features and identify the abnormal and normal snippets. The dual attention module is designed to be lightweight with a few numbers of trainable parameters. We also design a weighted loss to enhance the discriminative representation between the abnormal and normal segments. An extensive series of demonstrations conducted on large datasets such as UCF-Crime, ShanghaiTech and XD-violence shows that the proposed system has the potential to attain improved performance with an AUC of 84.98%, 95.32%, and 74.05%, respectively.