Weakly Supervised Temporal Attention Framework for Video Anomaly Detection
Manas Ranjan Biswal, Prabhu Prasad Dev, Santos Kumar Baliarsingh · 2024
Weakly supervised anomaly detection in surveillance videos is a formidable challenging task in visual understanding research community. A key factor contributing to the challenge of detecting abnormal events is the inability to effectively represent the temporal relationships between the snippets in videos. To address this challenge, our proposed approach incorporates a temporal attention module to model the temporal dependencies among video snippets. This module enables the network to focus on the more significant snippets by assigning attention scores to them. By multiplying the attention scores with the corresponding feature vectors, our method effectively detects anomaly events. Moreover, we explore the use of an inner-bag ranking loss function, which aids in learning discriminating representations and anomaly scores for efficient abnormal event detection. Experimental results obtained from the UCFC and LAD benchmark datasets validate the performance of the proposed method. It outperforms other existing methods, demonstrating superior performance in terms of anomaly detection accuracy and false alarm rate.