Deciphering the crowd: modeling and identification of pedestrian group motion

Area of Science:

• Computer Vision
• Social Signal Processing
• Pattern Recognition

Background:

• Associating attributes with pedestrians in crowds is crucial for applications like surveillance and customer profiling.
• Complex social settings and diverse attributes (e.g., relationships, hierarchy) complicate accurate pedestrian analysis.
• Existing methods struggle with the inherent complexity of social interactions within crowds.

Purpose of the Study:

• To develop a robust method for identifying social groups within pedestrian crowds.
• To leverage identified group structures for accurate pedestrian attribute association.
• To improve the understanding of social dynamics and relationships in crowded environments.

Main Methods:

• Exploitation of small group structures within crowds to infer social attributes.
• Identification of social groups using explicit motion models integrated via a hypothesis testing scheme.
• Development of two models (positional and directional relations) for a compound hypothesis testing scheme to determine group membership.

Main Results:

• Achieved high identification accuracy for social groups, ranging from 87% to 99% across three diverse datasets.
• Demonstrated the effectiveness of the proposed approach in various environmental properties and group characteristics.
• Successfully resolved ambiguities using a novel uncertainty measure based on local and global group relation indicators.

Conclusions:

• The proposed method effectively identifies social groups in crowds by analyzing relational models.
• Exploiting group structures provides reliable indicators for associating attributes with individuals.
• The approach offers a significant advancement in understanding and analyzing complex social dynamics in pedestrian crowds.