Path plan recognition has been a kind of online recognition using positions as inputs. To allow CGF to recognize opponents’paths and destinations in simulation, a recognition framework of Abstract Hidden Markov Model is introduced following analyzing the hierarchy of path plan. Since it is difficult to recognize the path plans using standard model when destinations are changed and plans are executed from top to bottom, the Abstract Hidden Markov Model with Changeable Top-level Policy is proposed. The initial distribution and termination variables of top policy were given and the relations between policy termination variables were adjusted to allow the lower policy for a forced termination. The modified DBN structure was presented, and the approximate inference was realized by deducing processes of updating conditional probability and sampling RB variables as well. Simulation experiments show that different kinds of typical paths in specific environment can be recognized efficiently with this method. The modified model not only confirms good recognition accuracy compared with the standard model under the circumstance when destination is not changing, but also performs well in solving destination changing path plan recognition problems with sufficient observation data provided.