A combined-observation pose correction method for enhancing tracking accuracy in visual measurement systems

Non-contact full-field 3D measurement and reconstruction using visual binocular tracking is crucial for many contemporary applications. However, accuracy limitations arise in long-range tracking due to low confidence in depth pose estimation. To address this, we propose a pose correction method using two binocular tracking units (TUs) in a perpendicular configuration. In the system, depth tracking from one unit is correlated to lateral tracking from the other unit through their relative pose. During fusion, lateral information takes precedence due to re-projection constraints, enabling the system to operate with high confidence in nearly any tracking direction. This approach not only improves tracking accuracy but also reduces spatial accuracy variations compared to conventional tracking systems. The proposed method is an optimization framework that leverages predictive models and corresponding loss functions. It estimates the combined tracking pose for each frame and the global relative poses between units using projection data from both TUs. Our system exhibits a maximum relative RMSE of 0.0074%, markedly lower than that of separate TU systems (0.0155–0.1153%). Furthermore, it achieves average reductions in spatial accuracy variation of 57.2% and 38.9% compared with individual TU systems.