With the recent advancement of 5G and B5G communication technologies, the scope of high- reliability application services utilizing drones-such as autonomous delivery, precision agriculture, infrastructure inspection, and swarm flight control-has rapi...
With the recent advancement of 5G and B5G communication technologies, the scope of high- reliability application services utilizing drones-such as autonomous delivery, precision agriculture, infrastructure inspection, and swarm flight control-has rapidly expanded. These emerging services fundamentally require high-precision real-time localization technology at the centimeter level to enable accurate takeoff, landing, and precision task execution. However, conventional localization methods reveal fundamental limitations in meeting such stringent precision requirements. The satellite-based navigation systems (GNSS) such as the Global Positioning System (GPS) suffer from severe performance degradation in dense urban environments or indoor spaces due to signal blockage and multipath errors caused by building reflections, which can reduce localization accuracy to several meters. Meanwhile, inertial navigation systems (INS) based on the Inertial Measurement Unit (IMU) have the advantage of operating independently of external signals, but inevitably accumulate drift errors due to the integration of noisy acceleration and angular velocity measurements, leading to a gradual degradation of positioning accuracy over time. To overcome these limitations, vision-based localization techniques have attracted significant attention. However, existing approaches still face practical challenges. For example, methods that estimate position by comparing images of the landing area with a pre-built database are limited in recognition range and often fail to achieve sufficient accuracy in distance and angle estimation, making precise drone landings difficult. Furthermore, when complex nonlinear optimization estimators are employed, the increased computational cost and processing delay hinder real-time control, thereby compromising system stability and responsiveness. Systems using multiple cameras can achieve higher precision, but they introduce significant complexity, requiring strict synchronization and calibration between cameras, and are also vulnerable to errors from lens distortion. To address these inefficiencies and instabilities while maintaining low cost and high reliability, this study proposes a novel integrated vision-based localization method that utilizes a single camera and a specially designed reference marker to accurately estimate the drone’s three-dimensional relative position and orientation (pose). The proposed method consists of three core processing stages, carefully designed to ensure both precision and real-time performance. First is high-precision camera calibration and lens distortion correction. Based on the pinhole camera model, the internal parameters of the lens and its radial and tangential distortion coefficients are accurately estimated. This process begins with an initial linear estimation using the Direct Linear Transform (DLT) method, followed by the application of the Levenberg–Marquardt algorithm, a nonlinear optimization technique that minimizes the reprojection error. This step effectively eliminates the fundamental causes of measurement errors during image acquisition, thereby enhancing the accuracy of subsequent estimation stages. Second is real-time three-dimensional position and pose estimation. Using the corrected intrinsic parameters and the predefined world coordinates of the reference marker, the system solves the Perspective-n-Point (PnP) problem in real time, establishing the correspondence between two- dimensional image coordinates and three-dimensional spatial coordinates. In this process, the extrinsic parameters-composed of a rotation matrix and a translation vector between the drone and the marker-are calculated, providing accurate information on the drone’s 3D position and orientation (Roll, Pitch, and Yaw), which enables precise autonomous control. Third is a window-based correction algorithm for angular stabilization. To effectively suppress outlier phenomena such as temporary spikes in Yaw, Pitch, and Roll values caused by gimbal lock during Euler angle conversion or sensor noise, a dynamic window-based stabilization method is employed. This technique stores recent angle data in a buffer, computes their average, and compares it with the current estimate to perform adaptive correction. This correction mechanism ensures smoother and more stable system operation, reducing errors caused by abrupt attitude changes. Experimental validation of the proposed integrated method under real-world conditions demonstrated that internal parameter calibration significantly reduced maximum measurement errors and ultimately achieved high localization accuracy. The results indicate that the proposed approach can serve as a key enabling technology for the future development of autonomous operation systems for drones and mobile robots by enhancing the precision of vision-based localization.