We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Dual-branch network with memory for video anomaly detection.
- Authors
Wang, Dicong; Hu, Qinghua; Wu, Kaijun
- Abstract
Anomaly event detection is a video surveillance technology automatically analyzing video sequences without manual intervention by employing machine learning and computer vision technology. In the existing approaches, most of them are utilized to reconstruct or predict the video frame based on an autoencoder (AE). However, impacted by the powerful characterization capabilities of Convolutional Neural Network (CNN), abnormal frames will be improperly reconstructed into normal frames. To solve the above issue, an autoencoder, based on a branch framework of reconstruction and prediction in training, is proposed. A memory module is adopted to reduce the reconstruction error, which is capable of enhancing the robustness of the autoencoder as a prototype memory module. The prediction of high-quality future frames can effectively prevent the reconstruction of abnormal frames, and the two branches can be supplemented with their respective loss functions, thus further elevating the performance of video anomaly detection. The framework for this study is trained from end to end. The methodology put forth in this article is extensively verified on three publicly available data sets, and its robustness to the uncertainty for the common occurrence as well as the efficiency to the sensitivity for the abnormalies are also confirmed.
- Subjects
INTRUSION detection systems (Computer security); ANOMALY detection (Computer security); VIDEO surveillance; COMPUTER vision; CONVOLUTIONAL neural networks; MEMORY; MACHINE learning
- Publication
Multimedia Systems, 2023, Vol 29, Issue 1, p247
- ISSN
0942-4962
- Publication type
Article
- DOI
10.1007/s00530-022-00991-x