Socially Constrained Structural Learning for Groups Detection in Crowd.

Area of Science:

• Computer Vision
• Social Network Analysis
• Computational Social Science

Background:

• Collective behavior emerges from interactions within small groups.
• Understanding crowd dynamics is crucial for various applications.
• Existing methods for group detection in crowds have limitations.

Purpose of the Study:

• To propose a novel algorithm for detecting social groups in crowds.
• To leverage people's trajectories for group identification.
• To develop a robust evaluation metric for group detection performance.

Main Methods:

• Utilizing a Correlation Clustering procedure on people's trajectories.
• Employing an online Structural Support Vector Machine (SVM) framework to learn affinities.
• Designing features inspired by Proxemic theory, Granger causality, Dynamic Time Warping (DTW), and Heat-maps to capture physical and social identity.
• Introducing a new loss function (G-MITRE) for evaluating group detection performance.

Main Results:

• The proposed algorithm achieves state-of-the-art results in social group detection.
• Performance is validated using both ground truth trajectories and tracklets from existing detector/tracker systems.
• The method effectively captures complex social interactions within crowds.

Conclusions:

• The developed algorithm provides a significant advancement in automated social group detection in crowds.
• The integration of trajectory data with advanced machine learning techniques offers a powerful approach.
• The G-MITRE loss function offers a more sociologically relevant evaluation of group detection.