Online anomaly detection in crowd scenes via structure analysis.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Behavioral Analysis

Background:

• Abnormal behavior detection in crowded scenes presents a significant challenge in computer vision.
• Existing methods often struggle with complex crowd dynamics and high densities.

Purpose of the Study:

• To develop a novel structure-based approach for detecting abnormal behavior in crowd scenes.
• To introduce an informative structural context descriptor (SCD) for individual crowd members.
• To enhance multi-object tracking for robust analysis in dense crowds.

Main Methods:

• Proposed an informative structural context descriptor (SCD) inspired by solid-state physics.
• Developed a robust multi-object tracker using 3-D discrete cosine transform (DCT) for SCD variation computation.
• Utilized online spatial-temporal analysis of SCD variations for abnormality localization.

Main Results:

• The proposed method effectively models relationships among individuals using SCD.
• The 3-D DCT multi-object tracker enables robust association of targets in high-density scenarios.
• Experimental results on public datasets demonstrate the method's effectiveness in detecting various anomalies.

Conclusions:

• The novel structure modeling approach offers a new direction for abnormal behavior detection.
• The SCD and 3-D DCT tracker combination provides a robust solution for complex crowd analysis.
• The method proves effective across diverse crowd video datasets.