Aiming at the safety and planning efficiency issues of autonomous mobile robots(AMR) path planning in complex unstructured three-dimensional environments, this paper proposed a three-dimensional path planning algorithm that combined improved Q-learning with improved dynamic window method(DWA) based on real terrain model data. Firstly, it used digital elevation model data(DEM) to evaluate the complexity of the ground terrain, and established a complex unstructured threedimensional environment model. Secondly, based on the three-dimensional environment model, Q-learning was improved by improving the reward and punishment mechanism and introducing the terrain feature initialization Q value table, and it adjusted the balance strategy of exploration and exploitation to accelerate the convergence speed. It designed the DIRRQ-learning three-dimensional global path planning algorithm to realize global path planning. On the basis of global path planning, the improved DWA algorithm was used to realize the dynamic real-time obstacle avoidance of AMR local path planning for unknown obstacles in the real-time three-dimensional environment model. Finally, the experimental results show that in terms of global path planning, the DIRRQ-learning algorithm improves path safety by 18.2% and convergence speed by 20.79% compared with the traditional Q-learning algorithm, in terms of local path planning, the improved DWA algorithm improves path safety by 2.4% compared with the traditional DWA algorithm. Therefore, the proposed three-dimensional path planning algorithm significantly enhances the path safety and planning efficiency in complex unstructured three-dimensional environments, improving the path planning capacity and efficiency of AMR.