Real-time multi-agent collision avoidance for large environments with hundreds or thousands of agents need powerful collision avoidance module. Most velocity-obstacles-based method for collision avoidance assume that every agent share the same responsibility to adjust their velocity to avoid potential collision. In order to improve the quality of dynamic collision avoidance for virtual pedestrian simulation, this study uses adjustment factor to distinguish the strategies of different type of pedestrian. And we introduce the less-effort to discuss the relationship between the velocity change and instantaneous energy consumption during dynamic collision avoidance. At last, we use linear programming to choose the best velocity from feasible velocity constructed by the improved ORCA algorithm. The experiment result shows that our method can improve the simulation efficiency of dynamic collision avoidance for large-scale crowd simulation, and the performance also meets the requirements of real-time simulation.