Robust Unsupervised Video Anomaly Detection by Multipath Frame Prediction.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video anomaly detection is crucial for security surveillance but remains challenging.
• Current deep reconstruction models often struggle with subtle differences between normal and abnormal frames.
• Frame prediction methods show promise but require refinement for real-world surveillance data.

Purpose of the Study:

• To propose a novel, robust, and unsupervised video anomaly detection method.
• To enhance frame prediction for better anomaly identification in surveillance videos.
• To address limitations of existing reconstruction-based approaches.

Main Methods:

• Developed a multipath ConvGRU-based frame prediction network to capture multi-scale spatial-temporal dependencies.
• Introduced a noise tolerance loss function to mitigate background noise interference during training.
• Employed an unsupervised learning strategy for anomaly detection.

Main Results:

• The proposed method achieved superior performance compared to state-of-the-art approaches on benchmark datasets (CUHK Avenue, ShanghaiTech Campus, UCSD Pedestrian).
• Achieved a frame-level Area Under the Receiver Operating Characteristic Curve (AUROC) score of 88.3% on the CUHK Avenue dataset.
• Demonstrated effectiveness in handling semantically informative objects and varying scales.

Conclusions:

• The proposed frame prediction-based method offers a robust solution for unsupervised video anomaly detection.
• The novel network architecture and loss function contribute to improved accuracy and noise resilience.
• The approach shows significant potential for real-world surveillance applications.