Autonomous landing on inclined and unstructured terrain remains a critical challenge for unmanned aerial vehicles (UAVs) in applications such as emergency rescue, infrastructure inspection, and disaster monitoring. Landing surfaces in these scenarios may be sloped, uneven, or discontinuous, considerably reducing stability and increasing the risk of propeller–ground contact. Although existing studies have explored adaptive landing gear and perception-based strategies, many rely primarily on simulation or limited fixed conditions for validation. Moreover, the integration of terrain perception, landing decision-making, and physical execution often remains insufficient for systematic experimental evaluation. A reproducible experimental platform that enables closed-loop verification of autonomous landing on unstructured terrain is therefore needed.

This study designs and builds a vision-based autonomous landing platform using a quadrotor UAV equipped with actively adjustable landing gear. The platform comprises modular components for terrain perception, landing attitude decision-making, execution control, and performance evaluation. A downward-facing depth camera captures depth images, from which point-cloud data are generated through projection using camera intrinsic parameters. To extract robust local terrain characteristics, the raw point cloud undergoes spatial filtering and region-based constraints before being segmented via plane fitting with the random sample consensus (RANSAC) algorithm. Principal component analysis (PCA) is then applied to the covariance matrix of each extracted plane cluster to estimate the surface normal vector, yielding the slope angle and slope aspect. Based on these terrain parameters, a geometric constraint model describes the relationship among UAV body attitude, landing-leg rotation angles, and local terrain configuration. This model incorporates mechanical constraints (servo angle limits) and the minimum airframe ground clearance, allowing collision avoidance constraints to be embedded directly in the decision-making process. The resulting landing planning problem is formulated as a constrained optimization task. The covariance matrix adaptation evolution strategy (CMA-ES) serves as the inner-loop optimizer, searching for feasible landing-leg angles and vehicle attitudes that minimize a weighted multi-objective cost function. To reduce manual tuning and improve robustness across varying terrain conditions, Bayesian optimization is introduced as an outer-loop strategy for adaptive hyperparameter tuning and accelerated convergence.

Experiments were conducted indoors to ensure safety and repeatability. Adjustable artificial slopes simulated representative unstructured landing conditions, including single-plane inclined surfaces and dual-plane configurations with discontinuities. A motion-capture system provided high-precision pose measurements for performance evaluation. Results demonstrate that the platform completes the full landing pipeline, from terrain perception and slope estimation through attitude optimization to controlled descent. Across repeated trials, the UAV achieved safe, stable landings on slopes up to 42°, with slope estimation errors within ±5° of the reference measurement. No propeller collision or landing failure occurred. Comparative analysis further shows that CMA-ES achieves faster convergence and more stable solutions than several baseline algorithms, while the outer-loop Bayesian tuning improves optimization consistency across different slope combinations.

This work presents a closed-loop experimental platform for autonomous UAV landing on unstructured terrain. By integrating depth-based terrain perception, geometric modeling with safety constraints, and optimization-driven attitude planning, the platform provides a reliable physical testbed for experimental validation. Its reproducible and extensible design supports rapid verification and iterative refinement of landing strategies and can serve as a practical foundation for research training and laboratory instruction on intelligent UAV landing in complex environments.