Data-driven methods, particularly Reinforcement Learning (RL), are pivotal for developing autonomous driving policies through interactive exploration and self-learning. However, RL approaches are often hindered by complex reward function engineering and a performance gap compared to human drivers. To address these challenges, this paper introduces a hierarchical Human-in-the-Loop RL (HITL-RL) framework that synergistically integrates real-time human decision feedback with multi-objective motion optimization. The framework comprises two core components: an upper-level, DDQN-based decision module that uses “dual human-prioritized stimulation terms” to reconstruct value distributions, accelerating alignment with human driving behavior; and a lower-level, cost-function-driven adaptive motion planner that concurrently optimizes driving efficiency, safety, and comfort based on environmental predictions. Simulation experiments demonstrate that the proposed HITL-RL strategy achieves faster and more stable convergence than a baseline RL method, enhances driving efficiency and safety—marked by a 20% increase in success rate—and shows alignment with human decision-making characteristics.