To address the challenges of reducing observation points and improving path quality under the coverage constraints of inspecting building complexes with the intelligent drone, this study proposes a two-stage planning framework integrating density-adaptive clustering and hierarchical energy consumption optimization. Firstly, at the observation point generation stage, a Density-Adaptive Clustering (DAC) algorithm is proposed based on the triangular mesh model of the building complex, and is further integrated with a Pareto-front-based multi-objective global optimization strategy considering the viewpoint quantity, coverage redundancy and observation quality, thereby enabling the efficient observation point generation. Additionally a multi-objective function was constructed for the path planning stage by jointly considering path length, turning angle, and elevation change. Based on this formulation, an improved Lin-Kernighan-Helsgaun algorithm incorporating a hierarchical energy consumption model and termed Stratification-LKH (S-LKH) was proposed to optimize the traversal sequence of observation points. To validate the superiority of the proposed method, the simulation experiments were conducted on the building models of two different scales. For the Lufu building complex, the number of viewpoints was reduced by 45.10%, 14.07%, and 27.78% compared to the voxel-expansion with random-key genetic algorithm, the two-stage optimization method, and the fuzzy clustering method, respectively. The proposed S-LKH algorithm also reduced the objective function values of 14.70%, 6.86%, 15.30%, 20.89%, and 13.82% compared with the LKH solver, variable-strategy reinforced algorithm, grey wolf-differential evolution hybrid algorithm, multi-strategy fusion differential evolution, and spherical vector-based particle swarm optimization, respectively. In terms of path features, the maximum turning angle and altitude change are decreased by up to 11.68% and 52.84%, while the average turning angle and altitude change are decreased by up to 10.08% and 22.82%. Finally, the simulation experiments and field flight tests further validate the effectiveness and engineering feasibility of the proposed method.