Small group human activity recognition

Small group people activity recognition has attracted much attention in computer vision community in recent years, since it has great potential in public security applications. Comparing to single human activity recognition, group human activity recognition has much more challenges, such as mutual occlusions between different people, the varying group size, and the interaction within or between groups. In this paper, we propose a novel structural feature set to represent group behavior as well as a probabilistic framework for group activity learning and recognition. We first apply a robust multiple targets tracking algorithm to track each individual in the entire image region. Small groups are then clustered based on the output positions of the tracker. After that, we introduce a set of social network analysis based structural features to describe the dynamic behavior of small group people in each frame. A Gaussian Process Dynamical Model(GPDM) is then employed to learn the temporal activity of small group people overtime. After training, the new group activity will be identified by computing the conditional probability with each learned GPDM. Our experimental results indicate that our proposed features and behavior model can successfully capture both the spatial and temporal dynamics of group people behavior, and correctly identify different group activities.