Design and Flight Validation of Monocular-vision-based Relative Altitude Estimation for Precision Landing of Lightweight Aerial Vehicles

Monocular-vision-based precision landing is promising for small unmanned aerial vehicles operating under strict payload constraints. However, detection noise and image disturbances can degrade altitude estimation accuracy and destabilize terminal landing performance. This study presents a vision-based precision landing framework that integrates a calibrated area - altitude model with an extended Kalman filter (EKF) to improve relative altitude estimation and landing accuracy. A You Only Look Once version 9 detector is used to extract the helipad center and bounding-box area, while the EKF employs a nonlinear area measurement model and an adaptive measurement covariance to suppress measurement uncertainty. The proposed framework was implemented in a robot operating system within a PX4 autopilot-Gazebo software-in-the-loop simulation environment and validated based on flight experiments using an X500 quadrotor platform. In actual flight tests, time of flight (TF) Mini Plus light detection and ranging (LiDAR) was used only as a reference sensor for performance evaluation. The EKF reduced the average altitude-estimation root-mean-square error from the LiDAR-referenced value to the EKF-estimated value. In addition, the proposed precision-landing algorithm reduced the landing-point error from the conventional return-to-home mode to the proposed mode, resulting in an improvement. The experimental results demonstrate that the proposed framework enhances the reliability of monocular vision-based altitude estimation and enables more accurate terminal precision landing than conventional return-based landing methods.