Automatic Recognition of Human Team Behaviors

Abstract

This paper describes a methodology for recording, representing, and recognizing team behaviors performed by players in an Unreal Tournament MOUT (Military Operations in Urban Terrain) scenario. To directly monitor the performance of human players, we developed a customized version of Unreal Tournament that records position and orientation of all the team members through time as they participate in a simulated MOUT scenario of a firing team moving through an urban area. Behavior recognition is performed offline using a set of hidden Markov models on short movement sequences that are translated into a canonical reference frame; the behavior model with the highest log likelihood for a given sequence is identified as correct. We believe that accurate offline recognition of team behaviors is an important prerequisite towards building virtual training environments for teamwork tasks.