MSADL: Video Anomaly Detection Based on Multi-Scale Attention and Dictionary Learning
Bo Wang, Hong Xia, Hui Jia, Yanping Chen · 2025
The goal of Video Anomaly Detection (VAD) is to identify anomalous events in videos that deviate from normal behavioral patterns. Due to the diverse behaviors and complex scenes of objects in surveillance videos, the application of this technique faces certain challenges. We propose a novel video anomaly detection method based on multi-scale attention and dictionary learning (MSADL) that can effectively integrate spatio-temporal information for accurate anomaly detection. unlike most existing methods, our approach is a unified architecture that is applicable to both unsupervised and weakly supervised VAD tasks. MSADL employs an efficient multi-scale attention mechanism to augment the temporal and spatial features of the image, and then employs the synergy between the dictionary-based representation and self-supervised learning to characterize the anomalies at the feature level to model anomaly concepts. Experimental results on two large-scale benchmark datasets, ShanghaiTech and UCF-Crime datasets, show that our approach outperforms the existing state-of-the-art and reaches a new optimal level of performance for both unsupervised and weakly supervised VAD tasks.