Lidar is widely used in the obstacle avoidance field of the ground mobile robot because of its high resolution of angle and range, but its inherent characteristics of detection may induce robots to be caught in the trap of V-obstacle easily, and increase the cost of planning path. Thus, the detection zone of lidar is divided into the indistinct avoidance region, the precise avoidance region and the emergent avoidance region respectively based on deep analysis of relationship between its accuracy and detection range, furthermore, a multilevel obstacle avoidance method is proposed. An obstacle avoidance algorithm based on neural networks is designed for the indistinct avoidance region, while in the precise avoidance region, collision avoidance is realized by tracking the boundary laser points. Simulation experiments verify that compared to traditional collision avoidance methods based on lidar, the proposed approach can not only make the robot avoid being caught in the trap of V-obstacle, but also generate the avoidance path at a smaller cost.