Learning Prompt-Enhanced Context Features for Weakly-Supervised Video Anomaly Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Weakly supervised video anomaly detection (WS-VAD) seeks abnormal activities in untrimmed videos without frame-level labels.
• Existing methods use graph convolutions or self-attention with multiple instance learning (MIL), but face high computational costs and limited intra-class discrimination.

Purpose of the Study:

• To develop a novel WS-VAD framework focusing on efficient temporal modeling and improved anomaly subclass discrimination.
• To address limitations of multi-branch architectures and binarized MIL constraints in prior works.

Main Methods:

• Introduced a Temporal Context Aggregation (TCA) module for efficient local-global dependency modeling using attention matrices and adaptive fusion.
• Proposed a Prompt-Enhanced Learning (PEL) module to integrate semantic priors via knowledge-based prompts for feature discrimination.

Main Results:

• The proposed WS-VAD framework demonstrated superior performance on UCF-Crime, XD-Violence, and ShanghaiTech datasets.
• Achieved better results with reduced parameters and computational effort compared to existing methods.
• Significantly improved detection accuracy for specific anomaly subclasses and lowered the false alarm rate.

Conclusions:

• The novel WS-VAD framework effectively enhances temporal modeling and intra-class anomaly discrimination.
• The TCA and PEL modules offer an efficient and effective solution for weakly supervised video anomaly detection.
• The approach shows promise for real-world applications requiring accurate and efficient anomaly identification.