A Feature-Trajectory-Smoothed High-Speed Model for Video Anomaly Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• High-speed detection of abnormal frames in surveillance videos is crucial for maintaining security.
• Existing methods may face challenges in achieving both speed and accuracy in anomaly detection.

Purpose of the Study:

• To propose a new video anomaly detection model, the Feature Trajectory-Smoothed Long Short-Term Memory (FTS-LSTM).
• To enhance the precision and speed of anomaly detection in surveillance videos.

Main Methods:

• The FTS-LSTM model utilizes an LSTM autoencoder to predict future frames from normal video streams.
• It employs a novel Feature Trajectory Smoothness (FTS) loss during training to improve temporal regularity learning.
• In the detection phase, both FTS loss and Generation Error (GE) loss are used as detectors.

Main Results:

• The FTS loss acts as a new indicator for anomaly detection, constraining the LSTM layer for precise temporal regularity learning.
• Cascading the FTS detector and GE detector enables high-speed and competitive anomaly detection performance.
• The model demonstrated effectiveness across multiple datasets.

Conclusions:

• The FTS-LSTM model offers a promising approach for efficient and accurate video anomaly detection.
• The integration of FTS loss significantly improves the model's ability to learn temporal dynamics.
• This method provides a robust solution for real-time surveillance security applications.