An Efficient and Robust Unsupervised Anomaly Detection Method Using Ensemble Random Projection in Surveillance...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Video anomaly detection is crucial for security and monitoring using surveillance cameras.
• Existing methods often involve computationally intensive back-propagation, limiting efficiency.
• Fully unsupervised learning is desired to avoid manual labeling of anomalies.

Purpose of the Study:

• To propose a novel, efficient, three-stage unsupervised method for video anomaly detection.
• To reduce the computational cost associated with traditional anomaly detection algorithms.
• To enhance the robustness and accuracy of unsupervised anomaly detection.

Main Methods:

• Utilized random projection in the first stage, replacing computationally expensive autoencoders.
• Formulated the optimization as a least-square regression problem with a closed-form solution.
• Employed a one-class support vector machine for normality estimation and an ensemble technique to combine detector scores for stability.

Main Results:

• The proposed method achieved superior performance compared to recent unsupervised anomaly detection techniques.
• The algorithm demonstrated effectiveness even surpassing some supervised approaches on benchmark datasets.
• Achieved comparable results to state-of-the-art unsupervised methods with significantly reduced running time.

Conclusions:

• The novel three-stage unsupervised approach is effective, efficient, and robust for video anomaly detection.
• The integration of random projection, least-square regression, and ensemble technology offers a significant advancement.
• This method provides a computationally feasible and high-performing solution for real-world anomaly detection applications.