Appearance-Motion United Auto-Encoder Framework for Video Anomaly Detection
Yang Liu, Jing Liu, Jieyu Lin, Mengyang Zhao, Liang Song
Abstract
The key to video anomaly detection is understanding the appearance and motion differences between normal and abnormal events. However, previous works either considered the characteristics of appearance or motion in isolation or treated them without distinction, making the model fail to exploit the unique characteristics of both. In this brief, we propose an appearance-motion united auto-encoder (AMAE) framework to jointly learn the prototypical spatial and temporal patterns of normal events. The AMAE framework includes a spatial auto-encoder to learn appearance normality, a temporal auto-encoder to learn motion normality, and a channel attention-based spatial-temporal decoder to fuse the spatial-temporal features. The experimental results on standard benchmarks demonstrate the validity of the united appearance-motion normality learning. The proposed AMAE framework outperforms the state-of-the-art methods with AUCs of 97.4%, 88.2%, and 73.6% on the UCSD Ped2, CUHK Avenue, and ShanghaiTech datasets, respectively.