In order to address the problems of insufficient consideration of terrain constraints, blind exploration in the early stage of reinforcement learning, and value overestimation in Q-learning for mobile robot path planning in three-dimensional unstructured terrain, a Double Q-learning path planning method integrating a D* heuristic value field and an adaptive search domain was proposed. First, a terrain traversability index was constructed from terrain data. The terrain traversability penalty, obstacle safety risk penalty, and maximum slope constraint were incorporated into the D* cost function. A global reference path, a D* heuristic value field, and a recommended action field were then generated. An adaptive search domain was constructed according to the D* reference path and local terrain complexity. The D* heuristic value field and recommended actions were incorporated into Q-table initialization, three-branch action selection, and reward function design. Finally, the planned path was smoothed. Simulation results show that the path lengths obtained by the proposed method on the two maps are generally better than those obtained by the comparison algorithms, and the maximum slopes remain below the safety threshold. In the ablation experiments, the cumulative terrain cost and the number of explored grid cells are reduced by 69.5% and 73.1%, respectively, compared with the baseline method without D* guidance and search-domain constraints. It is concluded that the proposed method can achieve a balance among path length, terrain traversal risk, slope constraints, and search efficiency.