This paper proposes an improved autonomous exploration method for complex environments. Firstly, based on SLAM, a hierarchical strategy is used to incrementally construct frontiers in a 3D occupancy map, and the mean-shift algorithm is used to cluster these frontiers to obtain candidate target points. Simultaneously, based on the input of the submap point cloud, the visibility graph is dynamically updated, and its vertices and edges are simplified. Secondly, an evaluation function that includes expected information gain and movement cost is adopted to select the target point. This function can better balance the relationship between information acquisition and cost consumption in different scenes through nonlinear adjustment. Furthermore, paths are planned on the visibility graph to guide the robot to explore the unknown environment quickly and avoid duplicate paths. Simulation experiment results show that compared with NBVP, our method reduces running time by 68% and travel distance by 38.6%, with the completion rate of NBVP being only 0.6. In addition, in the real-world scene, our method can also efficiently complete the exploration task. These results indicate that the algorithm effectively addresses the problems of overlooking local narrow areas and high path redundancy, thereby improving the efficiency of robot autonomous exploration.