Abstract
Autonomous driving technology is primarily developed to enhance traffic safety through advancements in motion prediction and adaptive control mechanisms. Highway lane merging remains a high-risk scenario, accounting for approximately 7% of highway collisions globally due to misjudged vehicle interactions, according to international statistics. This paper proposes a two-stage deep learning framework for autonomous lane merging in mixed traffic. Using the Argoverse dataset, which contains over 300,000 vehicle trajectories mapped to high-definition road networks, we first predict vehicle trajectories using a Seq2Seq model with LSTM layers, achieving a 21% improvement in prediction accuracy over a baseline Multi-layer Perceptron model. In the second stage, reinforcement learning is employed for maneuver generation, where a Dueling Deep Q-Network outperforms a standard DQN by 8% in collision avoidance. Experimental results indicate that the combined trajectory prediction and RL-based framework significantly reduces merging delays, enhances data-driven decision-making in mixed traffic environments, and provides a scalable solution for safer autonomous highway merging.