An unsupervised visual motion modeling method for aircraft takeoff and landing based on multi-scale temporal configurations and physical consistency constraints

The approach and landing phases involve frequent attitude adjustments and high sensitivity to external disturbances, making them among the most accident-prone flight segments. Conventional IMU/GPS-based attitude estimation is susceptible to drift and discontinuities under high-frequency vibration, structural coupling, and signal occlusion, which limits its reliability during complex touchdown maneuvers. To address the coupled motions observed in ground-based landing videos, including background translation, global aircraft attitude changes, and local structural vibrations, we propose an unsupervised sparse optical flow trajectory modeling framework. The proposed approach is based on multi-scale temporal configurations and physical consistency constraints. Sparse trajectories are constructed using Shi-Tomasi features and Lucas-Kanade optical flow with short- and long-term sliding windows. Motion descriptors capturing temporal stability, kinematic statistics, smoothness, and spatial distribution are extracted. In addition, a rigid-body physical residual derived from global translation estimation is introduced as a weak physical prior to improve interpretability and separability. Experiments on real civil airliner landing videos demonstrate that the proposed method can effectively isolate aircraft-related motion under complex background conditions. In scale aircraft flight tests with onboard IMU measurements, image-based rotational estimates exhibit strong trend-level consistency with IMU attitude signals. High correlation coefficients and physically interpretable temporal patterns are observed. These results demonstrate that the proposed approach can reliably extract physically meaningful motion representations under unlabeled and weak-prior conditions. The method provides a practical vision-based framework for attitude anomaly monitoring and health state modeling during the landing phase.