The confined spaces and complex terrain of underground environments pose significant challenges for the autonomous patrol and exploration by ground mobile robots. This paper proposes an autonomous navigation system designed for unstructured environments, utilizing terrain feature analysis. The system consists of three main stages. First, terrain cost is calculated by extracting terrain features—including roughness, slope, clearance rate, and curvature—from a global point cloud map while considering the robot’s kinematic constraints. Second, global path planning is conducted based on the computed terrain cost to generate a feasible route. Finally, motion primitives are generated offline through state space sampling, establishing a mapping between the primitives and the grid space. During navigation, real-time point cloud data, local terrain characteristics, and target locations are integrated to evaluate and rank candidate trajectories, enabling the selection of an optimal local path for execution. Simulation results illustrate the effectiveness and practicality of the proposed method.