Context Recovery and Knowledge Retrieval: A Novel Two-Stream Framework for Video Anomaly Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video anomaly detection seeks to identify deviations from expected behavior.
• Existing methods often rely on reconstruction or prediction errors, limited by local context and lacking a robust understanding of normality.
• These limitations hinder accurate detection of anomalous events in complex scenarios.

Purpose of the Study:

• To develop a more robust video anomaly detection method.
• To address the limitations of local context dependency in current approaches.
• To integrate both local context understanding and global normality knowledge for improved anomaly detection.

Main Methods:

• A novel two-stream framework combining context recovery and knowledge retrieval.
• Context recovery stream utilizes a spatiotemporal U-Net for future frame prediction with a maximum local error mechanism.
• Knowledge retrieval stream employs improved learnable locality-sensitive hashing (LSH) with Siamese networks and mutual difference loss to encode normality knowledge.

Main Results:

• The two-stream framework demonstrates effective complementarity between its components.
• Achieved state-of-the-art performance on benchmark datasets (ShanghaiTech, Avenue, Corridor) compared to methods without object detection.
• Showcased competitive or superior performance against methods utilizing object detection on ShanghaiTech, Avenue, and Ped2 datasets.

Conclusions:

• The proposed two-stream framework significantly enhances video anomaly detection capabilities.
• Integrating local context with learned normality knowledge provides a more comprehensive approach to identifying unusual events.
• The method offers a powerful and efficient solution for real-world video surveillance and analysis.