PROJECT TITLE :

Socially Constrained Structural Learning for Groups Detection in Crowd

ABSTRACT:

Trendy crowd theories agree that collective behavior is the results of the underlying interactions among small teams of individuals. During this work, we propose a novel algorithm for detecting social groups in crowds by means that of a Correlation Clustering procedure on people trajectories. The affinity between crowd members is learned through an on-line formulation of the Structural SVM framework and a collection of specifically designed options characterizing both their physical and social identity, impressed by Proxemic theory, Granger causality, DTW and Heat-maps. To adhere to sociological observations, we introduce a loss function ( -MITRE) able to house the complexity of evaluating group detection performances. We tend to show our algorithm achieves state-of-the-art results when wishing on each ground truth trajectories and tracklets previously extracted by accessible detector/tracker systems.