Bilayer sparse topic model for scene analysis in imbalanced surveillance videos

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Dynamic scene analysis is crucial for video surveillance applications.
• Detecting abnormal events in complex scenarios remains a significant challenge.
• Existing methods struggle with the imbalance between normal and abnormal activities.

Purpose of the Study:

• To mine semantic motion patterns from surveillance videos.
• To detect abnormalities that deviate from normal motion patterns.
• To develop a scene-independent approach for dynamic scene analysis.

Main Methods:

• Proposed a data-driven Bilayer Sparse Topic Model (BiSTM).
• Represented videos using a hierarchical generative process of words and documents.
• Treated motion patterns as latent topics and videos as mixtures of topics.
• Incorporated a one-class constraint to handle data imbalance and improve discriminative power.
• Employed an alternative iteration algorithm for model learning.

Main Results:

• The Bilayer Sparse Topic Model (BiSTM) effectively captures semantic motion patterns.
• The approach successfully detects abnormalities in complex dynamic scenarios.
• Experimental results on public datasets demonstrate the model's effectiveness and promise.

Conclusions:

• The proposed Bilayer Sparse Topic Model (BiSTM) offers a robust solution for dynamic scene analysis.
• The scene-independent and data-driven approach enhances the detection of abnormal events.
• The model's ability to handle imbalanced data makes it suitable for real-world surveillance.