Energy Level-Based Abnormal Crowd Behavior Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Behavioral Analysis

Background:

• Crowd behavior analysis is crucial for public safety and event management.
• Existing methods often struggle with camera perspective variations and timely anomaly detection.

Purpose of the Study:

• To develop a robust crowd abnormal detection method resilient to camera perspective effects.
• To accurately identify and timestamp crowd anomalies in real-time.

Main Methods:

• Pixels are treated as particles with velocities extracted via optical flow.
• Particle qualities are adjusted based on distance to mitigate camera perspective effects.
• Crowd motion segmentation and linear interpolation determine pedestrian distance and particle qualities.
• Co-occurrence matrix descriptors (consistency, entropy, contrast) analyze crowd behavior changes.

Main Results:

• The method effectively reduces camera perspective influence on crowd behavior analysis.
• Anomalous crowd behaviors are detected accurately and in a timely manner.
• Experimental results on the UMN dataset demonstrate the method's effectiveness.

Conclusions:

• The proposed energy-level distribution method offers a reliable approach for crowd abnormal detection.
• This technique enhances the accuracy and robustness of crowd behavior analysis systems.
• The method provides a timestamp for detected anomalies, aiding in incident response.