We utilize deep reinforcement learning to develop both single-agent and multi-agent methods that can accomplish autonomous drone surveillance tasks in a known indoor environment in this research. We combine the benefits of both visual and obstacle information to boost efficacy while ensuring low time consumption. And we devise a separate reinforcement learning training and test technique that both enhance training efficiency and ensure task completion. This method also creates a new field for sim-to-real transfer. Our experimental results show that the trained agents can detect all targets at a relatively fast speed while maintaining a high level of security, and the patrol completion rate is more than 98% in both single-agent and multi-agent tasks.