The crowd behavior recognition system is a subsystem of a distributed cooperative dynamic evacuation guidance system, and it detects and predicts crowd flow and abnormality occurrence by using machine learning method such as deep learning with visual and depth information obtained by RGB-D camera. The distributed cooperative dynamic evacuation guidance system aims to suggest evacuation routes at extensive evacuation situations by autonomously cooperation among plural sensors and evacuation guiding devices. In this paper, a recognition method of overall behavior of the crowd is proposed. Some indices for indicating the situation are examined in order to recognize the overall behavior of the crowd flow and the abnormality occurrence, and cooperate among the system by sharing the recognition results rapidly.