In view of the problem that the traditional guidance method is not accurate enough and is easily disturbed in complex sea conditions in the current autonomous landing of aircraft, in order to improve the landing accuracy and stability, this study proposes an autonomous landing guidance algorithm based on computer vision. The algorithm first uses the shipborne camera to collect aircraft images in real time; the improved YOLOv5 model is used for aircraft target detection, and the DeepSORT algorithm is combined to achieve stable target tracking. Based on the precise camera calibration parameters and the known three-dimensional model of the aircraft’s key points, the PnP algorithm is used to solve the spatial pose of the aircraft relative to the deck coordinate system. Finally, a data fusion module based on the extended Kalman filter is designed to fuse the visual pose estimation value with the aircraft motion state information provided by the inertial navigation to generate accurate guidance instructions such as lateral deviation, longitudinal deviation, sinking rate and heading deviation. In the simulated and real ship-swaying environment, the algorithm achieves a lateral position error of less than 0.15 m, a longitudinal position error of less than 0.2 m, an attitude angle estimation error of 0.65°, and a landing point positioning success rate of 97.8% during the entire guidance process from 1.5 km away from the landing point to the touchdown. The computer vision guidance algorithm significantly improves the accuracy, robustness, and reliability of autonomous landing of aircraft in complex sea conditions, and provides an effective solution for autonomous landing systems.