Semantic Features Extraction for Anomaly Detection from Real-World Videos*
Minh-Hanh Tran, Thanh-Hai Tran · 2023
Video-based anomaly detection opens up numerous practical applications such as video surveillance and healthcare. Despite having been studied for an extended period, this problem remains highly challenging because of variations in anomalies caused by people, objects, or contexts. Furthermore, the issue of missing annotated datasets and imbalances in the training data can complicate the development of automatic detection systems. Moreover, many existing systems have attempted to detect anomalies by performing binary classification without providing any explanation of the results. This paper proposes a method to detect anomalies from video while providing a semantic explanation of the cause. We first deploy off-the-shelf techniques to detect people and objects in the scene, then we extract semantic features (motion, appearance, posture) to recognize anomalies. To focus more on posture change, we introduce new postural descriptions and time series descriptions of poses using a spatiotemporal graph convolutional neural network. The proposed method is validated on three public datasets: UCSD Ped2, CUHK Avenue, and ShanghaiTech Campus showing improvements over the baseline methods.