An autonomous driving car-following algorithm based on optimized TD3 and its simulation

To address the issue of traffic congestion in urban highways, especially in narrow sections, this paper proposes a car-following strategy based on deep reinforcement learning. Specifically, this strategy attempts to optimize the traffic efficiency of autonomous driving while also ensuring passenger comfort. First and foremost, the paper develops a decision control model called Multi-Input Deep Fusion Twin Delayed Deep Deterministic Policy Gradient (MDFTD3), which is based on the TD3 algorithm. This model integrates autonomous vehicle kinematic data and visual information. By defining state and action spaces within the input space, the model minimizes the impact of road curvature changes compared to other vehicles and traffic interactions. The model outputs continuous control variables for vehicle driving actions, including acceleration, braking, and steering. Extensive high-precision highway simulation experiments in CARLA demonstrated that the MDFTD3 driving strategy improved traffic efficiency by 28.94% and passenger comfort by 62.88% compared with vehicles controlled by human. This study comes up with a new approach to alleviating traffic congestion in urban highways at narrow sections.