With the advancement of virtual broadcasting technology, the use of artificial intelligence animated characters in virtual scenes is becoming increasingly widespread. However, there are still a series of challenges and limitations to make the behavior of animated characters more natural, intelligent, and diverse. Therefore, this study proposes a behavior tree based animation character behavior modeling and a short-term memory action recognition method combining human geometric features. The research results indicate that when the behavior modeling model faces different obstacles, the successful avoidance rate is over 80%, and the avoidance reaction time is 0.41s-0.65s. The accuracy and loss function values of the action recognition method gradually converge to 1 and 0 with the quantity of iterations grows. For the recognition of seven types of actions, the accuracy of raising the left hand, raising the right hand, waving the left hand, and waving the right hand reaches 100%, and the recall rate of raising the right hand is 100%. The majority of action types have F-value scores above 0.9. Relative to the recurrent neural network model, the accuracy of the double-layer long-term and short-term memory model is 95.8%, which is significantly better than the former's 86.3%, showing better recognition performance. In summary, modeling and identifying the behavior of artificial intelligence animated characters can make the characters in virtual broadcasting more intelligent, natural, and realistic, thereby improving the viewing experience of virtual broadcasting, which has important practical value and research significance. This has significant practical and research value, providing insightful references for related fields.