Geometrically constrained position estimation through low-level tracking

Driven by the aerospace need to better monitor landing gear structures, we propose a vision-based position estimation framework to determine the configuration of a series of articulated rigid links in a harsh environment through leveraging knowledge of the subject geometry. The framework is capable of running on compute-limited hardware by exploiting the inherent efficiency of low-level point trackers. The framework is designed to simultaneously overcome localised image occlusions and scene vibration by combining a Kalman filter with a geometry-aware mixture model to realign tracking points. We demonstrate the effectiveness of compensating for low-cost tracker drift by presenting the position estimation performance under nominal and disturbance conditions and compare this approach to dedicated object-specific methods showcasing superior resilience to typical disturbances and high estimation speed.