Abstract: To improve the interaction of virtual pedestrians who have been integrated into a real scene video, we present an intelligent approach to dynamically adjust the path of the virtual pedestrians according to the distribution of real crowd in the video.It first set up a hierarchical model of the scene allowing the virtual pedestrians to perform path planning at various levels.During the simulation, the future positions of real pedestrians are predicted and the velocity and density of real crowd are recorded at the respective scene grids.A detection ray is then shot along the way of the virtual pedestrians to check the crowd distribution in the front and the detection range conforms with current velocity of the virtual pedestrians.Experiment results based on several test scenarios demonstrate that the proposed approach allows better performance of virtual pedestrians when dealing with the real crowd compared with that by previous approaches.