Enhancing Video Anomaly Detection Using Spatio-Temporal Autoencoders and Convolutional LSTM Networks

• Published in: SN computer science Vol. 5; no. 1; p. 190
• Main Authors: Almahadin, Ghayth, Subburaj, Maheswari, Hiari, Mohammad, Sathasivam Singaram, Saranya, Kolla, Bhanu Prakash, Dadheech, Pankaj, Vibhute, Amol D., Sengan, Sudhakar
• Format: Journal Article
• Language: English
• Published: Singapore Springer Nature Singapore 11.01.2024; Springer Nature B.V
• Subjects: Algorithms; Anomalies; Automation; Behavior; Cameras; Computer Imaging; Computer Science; Computer Systems Organization and Communication Networks; Data Structures and Information Theory; Datasets; Deep learning; Frames (data processing); Information Systems and Communication Service; Localization; Original Research; Pattern recognition; Pattern Recognition and Graphics; Pedestrians; Soft Computing in Engineering Applications; Software Engineering/Programming and Operating Systems; Spatial data; Supervised learning; Surveillance; Video data; Vision; Walkways; Suspicious activities; LSTM; Spatio-temporal features; Autoencoders; Anomaly detection; Machine learning
• ISSN: 2661-8907; 2662-995X; 2661-8907
• DOI: 10.1007/s42979-023-02542-1

Identifying suspicious activities or behaviors is essential in the domain of Anomaly Detection (AD). In crowded scenes, the presence of inter-object occlusions often complicates the detection of such behaviors. Therefore, developing a robust method capable of accurately detecting and locating anomalous activities within video sequences becomes crucial, especially in densely populated environments. This research initiative aims to address this challenge by proposing a novel approach focusing on AD behaviors in crowded settings. By leveraging a spatio-temporal method, the proposed approach harnesses the power of both spatial and temporal dimensions. This enables the method to effectively capture and analyze the intricate motion patterns and spatial information embedded within the continuous frames of video data. The objective is to create a comprehensive model that can efficiently detect and precisely locate anomalies within complex video sequences, specifically those featuring human crowds. The efficacy of the proposed model will be rigorously evaluated using a benchmark dataset encompassing diverse scenarios involving crowded environments. The dataset is designed to simulate real-world conditions where millions of video footage need to be continuously monitored in real time. The focus is on identifying anomalies, which might occur within short time frames, sometimes as brief as five minutes or even less. Given the challenges posed by the massive volume of data and the requirement for rapid AD, the research emphasizes the limitations of traditional Supervised Learning (SL) methods in this context.