Reinforcement learning (RL) has been widely adopted to make intelligent driving policy in autonomous driving due to the self-evolution ability and humanoid learning paradigm. Despite many elegant demonstrations of RL-enabled decision-making, current research mainly focuses on the pure vehicle driving environment while ignoring other traffic participants like bicycles and pedestrians. For urban roads, the interaction of mixed traffic flows leads to a quite dynamic and complex relationship, which poses great difficulty to learn a safe and intelligent policy. This paper proposes the encoding integrated decision and control (E-IDC) to handle complicated driving tasks with mixed traffic flows, which composes of an encoding function to construct driving states, a value function to choose the optimal path as well as a policy function to output the control command of ego vehicle. Specially, the encoding function is capable of dealing with different types and variant number of traffic participants and extracting features from original driving observation. Next, we design the training principle for the functions of E-IDC with RL algorithms by adding the gradient-based update rules and refine the safety constraints concerning the otherness of different participants. The verification is conducted on the intersection scenario with mixed traffic flows and result shows that E-IDC can enhance the driving performance, including the tracking performance and safety constraint requirements with a large margin. The online application indicates that E-IDC can realize efficient and smooth driving in the complex intersection, guaranteeing the intelligence and safety simultaneously.