Abstract: In agent-based crowd simulation,previous methods for computing orientation will lead to abrupt changes in agent’s orientation,or generate stiffly turning behaviors.Therefore,the 3D performance of high-density crowd simulation deviates from reality.This paper presents a double-layer orientation fairing method for agent-based crowd animation.The theoretical layer computes an initial value of the individual orientation,considering the expected direction,the angular momentum constraints and the interfered rotation caused by surrounding obstacles.The numerical layer uses filtering technology for processing orientation data.The final orientation is interpolated between results from the two layers in hyperspherical space of quaternion.The method has proved to be effective to generate non-oscillatory and natural individual orientation,even for high-density crowd in complex environment.