With the growing need for drones in complex and hazardous environments, this paper proposes a robust autonomous exploration method to address key limitations in existing algorithms, such as blind viewpoint selection, path redundancy, and excessive map nodes. Locally, a hybrid viewpoint generation strategy based on boundary geometry and space filling is introduced, along with a gain evaluation mechanism tailored to viewpoint types for improved path targeting. Globally, a dynamic viewpoint graph is constructed via branch extraction, with node pruning and edge constraints guided by historical and planning paths. A visibility-aware global path optimization further enhances smoothness and safety. Experiments show the proposed method reduces exploration time by approximately 12%, path length by approximately 21%, and improves efficiency by approximately 23% over baseline methods.