A distributed algorithm is proposed to solve the problem of unmanned surface vessels (USV) avoiding collision with other normal and abnormal USVs based on the improved dynamic window approach (DWA). The algorithm also includes collision risk detection and responsibility division for driving. Firstly, prediction trajectories of obstacles and weighting factors are utilized to improve the distance evaluation function of the traditional DWA, which can avoid multiple dynamic obstacles better. While a novel rule evaluation function considering international regulations for preventing collisions at sea (COLREGS) is introduced to constrain actions of USVs. Secondly, an existing collision risk detection method is modified by adding desired velocity and course to reduce the fluctuation caused by changes of collision risk. Then, we propose a responsibility division method for multiple USVs to handle the problem that COLREGS only specify the method for two ships encounter, which takes abnormal USVs into consideration. Finally, simulations of multiple USVs collision avoidance are realized based on Matlab. The results show that USVs utilizing the distributed collision avoidance algorithm can make safe actions in accordance with COLREGS even if there are abnormal USVs.