A novel spatio-temporal memory network for video anomaly detection

Future frame prediction for anomaly detection methods based on memory networks have been extensively explored in the academic domain. Nevertheless, traditional memory-guided network techniques, which store dispersed spatial low-dimensional features, often fall short in delivering satisfactory results when applied to datasets characterized by variable scenes. This deficiency is evident in the frequent challenges faced during network convergence in the training process, resulting in unstable training outcomes. In response to this challenge, we introduce a novel Spatio-Temporal Memory Module, denoted as ST_MemAE. Our approach is designed to retain temporal correlation information within low-dimensional features, enhancing the representation of temporally closely linked features within the output of the encoder. Furthermore, we incorporate a homogeneous uncertainty function to optimize the balance of weights associated with multiple loss functions that are part of the memory module update process. As a result, our method offers improved stability in model training, faster convergence, and higher quality predictions of future frames. To validate the effectiveness of our approach, we conducted extensive experiments utilizing three distinct video anomaly detection datasets: UCSD Pedestrian 2, CUHK Avenue, and ShanghaiTech. The outcomes of these comprehensive experiments on publicly available datasets underscore the robustness of our method in accommodating diverse normal events while maintaining sensitivity to abnormal events. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.