A data stream-based approach for anomaly detection in surveillance videos

• Published in: Multimedia tools and applications Vol. 83; no. 21; pp. 60213 - 60241
• Main Authors: Aydogdu, Ozge, Ekinci, Murat
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.06.2024; Springer Nature B.V
• Subjects: Algorithms; Anomalies; Artificial neural networks; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Data transmission; Deep learning; Ensemble learning; Machine learning; Multimedia Information Systems; Segments; Special Purpose and Application-Based Systems; Surveillance; Video; Video surveillance; Data stream; Incremental learning; Concept drift; Anomaly detection; UFROS-ELM
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-17861-x

Anomaly detection is a challenging task in surveillance videos. Nowadays, deep learning-based approaches have been developed for this issue. Although they achieve good performance, the training procedure is time-consuming and expensive due to including many layers. Besides, they have a dependence on the training set, and the structure of the deep networks is not suitable for dynamic events. They require retraining to detect different normal or abnormal events. They cannot adapt to the new event dynamically and instantly without retraining. This paper aims to bring a novel perspective to anomaly detection in surveillance video by tackling the task in a data stream manner to overcome these disadvantages. A novel and simple data stream-based ensemble approach for video anomaly detection is presented in this paper. Initially, fixed-sized temporal segments are created using current frames during streaming. A multiple instance learning-based preprocessing method is applied to the current segment, and a 1-D flow vector is obtained. A fixed-sized chunk is generated by pooling the flow vectors for learning. Afterwards, a streaming data learning algorithm called Unsupervised Feature Representative Online Sequential-Extreme Learning Machines (UFROS-ELM) is applied to the current chunk. UFROS-ELM makes the initial prediction about the current vectors using a concept drift detection mechanism and ELM-based autoencoder. Finally, multiple UFROS-ELM based ensemble learning is employed for the final decision using the majority voting. The results are achieved on the well-known surveillance data sets and compared with state-of-the-art deep learning-based video anomaly detection algorithms. The promising results support further research in this area.