Temporal-Enhanced and Visual-Text Adaptive Fusion for Weakly Supervised Video Anomaly Detection in Public Safety

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Manual monitoring of surveillance video is inefficient and prone to errors.
• Automated anomaly detection is crucial for intelligent security systems.
• Weakly supervised video anomaly detection (WS-VAD) presents unique challenges.

Purpose of the Study:

• To propose a robust WS-VAD model for improved threat identification.
• To enhance the accuracy and efficiency of security systems.
• To address the limitations of current video anomaly detection methods.

Main Methods:

• Developed the Temporal-Enhanced and Visual-Text Adaptive Fusion (TE-VTAF) model.
• Introduced a Dynamic Local-Global Temporal Adaptive Module (DLG-TAM) for temporal dependency extraction.
• Implemented a Visual-Text Adaptive Fusion Module (VTAFM) for cross-modal feature aggregation.
• Utilized Multiple Instance Learning (MIL) with novel loss functions (Top-K outer bag and K-maxmin inner bag loss).

Main Results:

• The TE-VTAF model demonstrated superior performance in WS-VAD.
• Achieved an AUC of 88.93% on the UCF-Crime dataset.
• Achieved an AP of 85.62% on the XD-Violence dataset.
• Outperformed existing state-of-the-art methods on large-scale benchmarks.

Conclusions:

• The proposed TE-VTAF model offers a robust solution for weakly supervised video anomaly detection.
• The model effectively captures temporal dynamics and fuses visual-textual information.
• This advancement contributes to more reliable and efficient intelligent security systems.