This paper is concerned with the performance of different state representation on the vehicle decision-making problem at unsignalized intersection based on deep reinforcement learning. A hybrid state representation architecture based on attention mechanism and collision time prediction is designed, which can effectively improve the autonomous decision-making ability of vehicles. Compared with the traditional state representation method, the feature of our method is that the fusion of features can improve the vehicle’s obstacle avoidance ability, and the introduced action masking module can improve the vehicle’s traffic efficiency. Finally, test results in the simulation environment verify the effectiveness of our proposed method.