Abstract: To address the challenge of safe navigation for unmanned autonomous vehicles in highly dynamic, obstacle-dense unstructured environments, a hierarchical path planning framework is proposed to improve the efficiency, smoothness, and feasibility in planning. In the global planning layer, an RRT* (rapidly-exploring random tree star) method incorporating maximum vehicle turning angle constraints is developed. A dynamic KD-tree is adopted to accelerate nearest-neighbor search, and B-spline curves are combined to ensure path smoothness, producing a curvature-continuous global reference path for unmanned vehicles in unstructured cluttered environments. In the local planning layer, an adaptive weighted artificial potential field is introduced to dynamically adjust goal-directed and environmental potential weights, enhancing obstacle avoidance and stability. Guided by the global path, the planner respects curvature constraints and enables rapid response to both dynamic and static obstacles. Comparative experiments in typical off-road, dynamic mixed-obstacle, and campus environments using grid maps and ROS (robot opertating system) show that, compared to existing sampling-based methods, the proposed approach reduces average planning time by 18%~20%, node count by 30%~36%, path length by 8%~12%, and significantly improves obstacle avoidance success rate, generating smooth trajectories satisfying vehicle kinematic constraints in complex dynamic environments. Overall, the proposed hierarchical path planning framework offers real-time performance, robustness, and practicability in complex dynamic environments, providing an effective and reliable solution for autonomous navigation of unmanned vehicles in unstructured scenarios.