The invention discloses a group behavior identification method based on multi-feature fusion. The method includes extracting feature information of three different levels, i.e., extracting features of single persons, which mainly considers the position, size and motion information of each person in each frame and uses a convolutional neural network to extract the features of each person, extracting semantic features aiming at interpersonal interaction, which mainly considers an interpersonal external action relation and relative orientation relation, and extracting scene information on environment where people is in group behavior. A full-link conditional random field model is adopted, and the feature information is fused, thereby realizing identification of group behavior. The method considers various feature information at the same time, can describe group behavior more comprehensively and effectively, improves the identification rate of group behavior, and has important application value in video monitoring.