Anomaly Detection Based on Cascaded Swin Transformer
Yaqoob Raffay, Limin Xia, Syed Akram · 2024
Abnormal behavior detection in surveillance aims to identify actions that deviate from expected patterns, highlighting potential security threats or safety issues. Its significance lies in enabling early detection and response to incidents, thus enhancing public safety and security. Despite progress, existing algorithms often struggle with adapting to diverse environments and ensuring real-time processing. This research introduces a novel approach using a Swin-LSTM-based model that leverages the spatial analysis capabilities of Swin Transformers and the temporal insights of LSTM networks. To facilitate real-time application, the model complexity is minimized through Singular Value Decomposition (SVD) on the attention matrix and weight pruning within the MLP layers. Demonstrating superior performance on standard datasets, this method represents a significant step forward in efficient, real-time detection of abnormal behavior.