Abstract: In this paper, a method for video events learning and recognition based-on key atomic actions is presented. First, the hierarchical structures of events and the temporal relations between sub-events and atomic actions are expressed by and-or graph, which is learned from training data by the minimum description length principle. Then, the weights of the atomic actions are computed according to their importance, and the atomic action with the maximum weight is regarded as the key atomic action. The weight values of the atomic actions could be used for real-time events parsing which leads to the improvement of event recognition ratio. We also define the event recognition degree based on the weight of atomic action to reduce the number of identified events and raise the efficiency of events recognition algorithm. Finally, the experimental results show that, in a variety of scenes, the proposed method is effective for event recognition.