Robust GPS/Visual–Inertial Navigation in Dynamic and GPS-Degraded Environments via Dynamic Feature Masking and GPS Measurement Screening

Achieving reliable localization in dynamic environments remains challenging because moving objects corrupt visual constraints, and degraded GPS measurements introduce erroneous absolute-position corrections. To address these challenges, this paper presents a robust GPS/visual–inertial navigation (VIO) framework that enhances the reliability of visual and GPS measurements at the measurement level without modifying the underlying VIO estimator. In the visual front end, a YOLOv8-based detector asynchronously processes downsampled images to identify potentially dynamic regions, which are subsequently mapped back to the original resolution to generate a dynamic mask. Features detected within these masked regions are excluded from subsequent visual state estimation. For GPS integration, receiver-provided latitude, longitude, and altitude (LLA) coordinates are first transformed into a local East-North-Up (ENU) frame. Temporally aligned initial segments of the GPS and VIO trajectories are then used to estimate the relative yaw rotation between the two coordinate frames, thereby expressing GPS positions directly within the VIO frame. Prior to each GPS update, a two-stage screening strategy combining receiver-reported covariance with normalized innovation squared (NIS) testing rejects position measurements exhibiting excessive uncertainty or statistical inconsistency with the VIO state prediction. Vehicle field experiments in a dynamic-traffic environment demonstrate that dynamic-feature masking reduces the horizontal RMSE from 2.074 m to 1.820 m and the maximum horizontal error from 17.823 m to 11.963 m. In a GPS-degraded urban-canyon scenario, the proposed screening strategy reduces trajectory jitter from 0.0425 m to 0.0218 m while mitigating positional and kinematic disturbances, as indicated by the trajectory-jitter and kinematic metrics. These results demonstrate that measurement-level screening improves GPS/VIO localization robustness under the evaluated dynamic-traffic and GPS-degraded conditions. The current validation focuses on ground-vehicle operation in urban traffic and GPS-degraded environments; the effectiveness of the asynchronous masking strategy under highly maneuvering platforms and rapidly changing object motion requires further investigation.