When the mobile robot passes through the dynamic dense crowd, due to the insufficient understanding of environmental information, the robot navigation efficiency is low and the generalization ability is weak. To solve this problem, a double attention deep reinforcement learning algorithm is proposed. Firstly, the sparse reward function is optimized, and the distance penalty term and comfort distance are introduced to ensure that the robot approaches the target while taking into account the safety of navigation; Secondly, a state value network based on double attention is designed to process environmental information to ensure that the robot navigation system has both environmental understanding ability and real-time decision-making ability; Finally, the algorithm is verified in the simulation environment. Experimental results show that the proposed algorithm not only improves the navigation efficiency, but also improves the robustness of the robot navigation system. The main performance is that in 500 random test scenarios, the collision times and timeout times are 0, the navigation success rate is better than the comparison algorithm, and the average navigation time is 2% shorter than the best algorithm; When the number of pedestrians and navigation distance in the environment change, the algorithm is still effective, and the navigation time is shorter than the comparison algorithm.