AFRL-AFOSR-JP-TR-2016-0044 Activity Recognition in Social Media

In this work, we present a novel approach to analyze crowd behavior at various levels of granularity − individual, group and global. We first model the collective motion of the agents present in the scene by a first order dynamical system. The model learns the spatio-temporal interaction pattern of the crowd which is further analyzed for group detection. The groups are identifiable from the eigenvectors of the interaction matrix of the model and can be recovered by employing a variant of spectral clustering on the eigenvectors. We show that while eigenvectors detect groups, the eigenvalues characterize various group activities such as stationary, walking, splitting and approaching. Finally we classify a crowd video in one of the eight categories by employing a random forest. As an application, the model is used to predict personal space violation.