Urban air mobility (UAM) requires safe, precise takeoffs and landings at vertiports and helipads. However, in urban environments, global navigation satellite system (GNSS)-based positioning accuracy can be degraded by signal blockage and multipath propagation, reducing landing precision. Because helipads are designed for a predefined approach direction, precision landing also requires accurate position and heading alignment. This paper proposes a vision-based precision landing system for UAM operations in GNSS-denied environments using a You Only Look Once version 11 (YOLOv11)-oriented bounding box model trained with a convolutional neural network and a linear Kalman filter (LKF). The proposed system performs GNSS-independent position estimation using an optical flow sensor while estimating the helipad’s relative position and heading from images acquired by a downward-facing camera. Vision-based measurements and LiDAR altitude information are fused through the LKF to estimate the vehicle’s state. Based on the estimated relative position and heading, control and heading alignment are performed during precision landing. The proposed algorithm was implemented in a robot operating system within a PX4 autopilot environment and validated through both Gazebo-based Software-in-the-Loop simulations and actual flight tests. Flight tests were conducted in three scenarios: GNSS-based Return-to-Home (RTH) mode, use of the proposed precision landing algorithm in a GNSS environment, and use of the proposed precision landing algorithm in a GNSS-denied environment. Experimental results demonstrate that the proposed algorithm achieved stable state estimation and precision landing in both GNSS and GNSS-denied environments, while providing improved landing accuracy and successful heading alignment compared with the conventional GNSS-based RTH mode. These results verify the effectiveness and practical applicability of the proposed precision landing system in GNSS-denied environments.