Hierarchical group process representation in multi-agent activity recognition

In this paper, we develop a novel multi-agent activity recognition method, which emphasizes the influences of group on individuals and the hierarchical dynamics of group activities. We believe that the dynamics of group process plays a dominative role in characterizing multi-agent activities, and interactive information between agents is embedded in the influences of the group on individuals. Within the dynamic probabilistic network framework, we present a hierarchical control model (HCM) which consists of three parts: multi-channel layer, control layer and abstract layers. The multi-channel layer models the individual dynamics, at which agent channels are independent conditioning on the control layer. HCM extracts the group process using the control layer and represents its hierarchical characteristics with multi-level abstract layers. We combine multiple feature streams coming from individuals and recover the complex structure of the activity in one compact model. In experiments, the performance of HCM is evaluated and is compared with some other models. The results show that HCM is suitable for recognizing long complex multi-agent activities.