Abstract: Traditional studies on crowd simulation usually regard crowds as the combinations of individuals,but in fact,crowds consist of different small groups of individuals with similar behaviors.Most of the traditional crowd simulation methods ignored this property,and thus,produced unreal results.To overcome this problem,after considering the fact that a crowd consists of different small groups,this paper proposed a novel visually-guided approach containing both collision avoidance and dynamic navigation for simulating crowds.The simulation results show natural-looking self-organized phenomena which are more obvious and appealing in high density or complex situations compared with existing classical models.