Video Anomaly Detection with Sparse Coding Inspired Deep Neural Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Sparse coding has shown success in anomaly detection.
• Existing methods may have high computational costs and require careful hyperparameter selection.

Purpose of the Study:

• To develop an efficient and effective anomaly detection method.
• To improve upon sparse coding-based anomaly detection using deep neural networks.

Main Methods:

• Proposed Temporally-coherent Sparse Coding (TSC) optimized via Sequential Iterative Soft-Thresholding Algorithm (SIATA), equivalent to stacked Recurrent Neural Networks (sRNN).
• Developed an sRNN-Autoencoder (sRNN-AE) by adding a reconstruction layer to the sRNN.
• Introduced data-dependent similarity learning, reduced sRNN depth for real-time inference, and employed temporal pooling for enhanced robustness.

Main Results:

• The sRNN-AE method significantly outperforms existing anomaly detection techniques.
• Achieved real-time anomaly detection capabilities.
• Demonstrated effectiveness on both controlled and real-world datasets, validated by a newly created large-scale dataset.

Conclusions:

• The proposed sRNN-AE method is effective for anomaly detection.
• The approach offers real-time performance and improved robustness.
• The developed large-scale dataset facilitates future research in anomaly detection.