Tightly Integrated Smartphone GNSS and Visual odometry for Enhanced Urban Pedestrian Positioning

Precise smartphone-based positioning service is challenging in dense urban areas due to significant multipath effects in GNSS signals received by smartphone devices. The raw GNSS measurements will be contaminated by non-line-of-sight (NLOS) signals, severely deteriorating the smartphone positioning accuracy. Many methods have been proposed to mitigate the GNSS NLOS problem, including 3D mapping-aided GNSS, RAIM, and machine learning-based methods. But these methods have limitations such as the need for 3D city models or external devices, high false-alarm chances, and training processes. In this study, we have developed a new approach to improve smartphone positioning accuracy in dense urban areas by coupling the smartphone GNSS and camera sensors, which are already available in most smartphones. Wholly based on themselves, the proposed method tightly integrates GNSS pseudorange, carrier-phase and Doppler measurements, and a visual odometry (VO). The GNSS measurements undergo preprocessing, DD normal equations, and velocity estimations. The smartphone images are processed using a KLT optical flow method, where GNSS velocities are applied to estimate the coordinate rotation and scale between them based on a sliding-window least-squares scheme using Horn’s method. Importantly, a quad-tree-based outlier searching (QTOS) algorithm is applied to ensure the healthiness of estimation processes throughout the integration. The data from DD GNSS normal equations, GNSS velocities, and VO velocities are input to an FGO algorithm for final positioning estimations. A field test in the dense urban area of Calgary showed an improvement of 25% in horizontal accuracy and a reduction of velocity estimation error by 30%, where the chance of positioning outliers (> 30 m) is significantly reduced by 76%. Therefore, the proposed method provides an effective solution for precise smartphone positioning in dense urban areas without the need for external data sources or training.