Empirical Test of Visual Recognition System of Autonomous Vehicle Based On Deep Learning

The paper describes the development of an automatic driving vehicle visual recognition system leveraging deep learning models. It is based on integrated deep reinforcement learning and Markov decision process theory. The foundational model of this system utilizes a standard voting method, combining three basic network models: Deep Q-Learning Network (DQN), Double DQN (DDQN), and Dueling Double DQN (Dueling DDQN). Tests and generalization validation were conducted in a high-speed highway simulation environment, considering various lane configurations and driving behaviors such as lane changes to the left, lane keeping, lane changes to the right, lane maintaining acceleration, and deceleration. Results indicate that compared to other models, this integrated model achieved respective increases in decision success rates by 6%, 3%, and 6%. Additionally, there was an improvement in average vehicle speed, and decision-making time was less than 1 millisecond over 100 rounds of testing, meeting real-time requirements. This automatic driving vehicle visual recognition system effectively achieves autonomous driving functionality.