Kinematically Constrained Joint Pose Estimation of a UAS and Non-Cooperative Moving Platform

This work presents a Kinematically-Constrained VIO (KC-VIO) framework for jointly estimating the pose of an Uncrewed Aerial System (UAS) and a non-cooperative moving platform in GPS-denied environments. While standard Visual-Inertial Odometry (VIO) approaches model targets as unconstrained point masses governed by Random Walk (RW) processes, these models fail to capture physical motion constraints, leading to inconsistent state estimates and heading drift. To address this limitation, we fuse stereo vision and inertial measurements within an Extended Kalman Filter (EKF) that incorporates a kinematic bicycle model of the platform. By introducing a novel steering-angle state, we explicitly enforce non-holonomic constraints (zero side-slip) and couple the target’s linear and angular velocities, effectively constraining the evolution of the relative heading without requiring external global measurements. The formulation leverages both stationary and platform-based visual features and accounts for the cross-covariance between UAS and platform errors. Evaluation using real-world flight data demonstrates that, compared to standard Random Walk and simple Non-Holonomic Constraint models, the proposed kinematic formulation significantly reduces relative heading error and improves estimator consistency.