Robust and continuous localization is essential for autonomous driving but remains challenging in urban canyons where Global Navigation Satellite System (GNSS) signals are frequently blocked or degraded. This paper proposes a reliability-adaptive fusion framework that integrates commercial GNSS/inertial navigation system (INS) modules and visual–inertial odometry (VIO) using an interacting multiple model (IMM) filter. The proposed method employs a role-sharing sequential correction structure: VIO provides only relative motion information (velocity and trajectory shape), while GNSS/INS serves as the absolute position anchor to eliminate accumulated drift. Within each IMM mode, VIO-based motion correction is performed first, followed sequentially by GNSS-based position correction. The residuals from both correction stages are then used to update the mode probabilities of the underlying vehicle maneuver models, i.e., constant acceleration and constant turn. Furthermore, the dynamic covariance output of each estimator is treated as a real-time reliability indicator, enabling adaptive fusion weighting without manual threshold tuning. The proposed framework is validated on both a public benchmark (KAIST Urban Dataset) and a custom dataset collected with an instrumented vehicle equipped with commercial off-the-shelf sensors. The results demonstrate consistent improvements in positioning accuracy and trajectory continuity across diverse GNSS-denied scenarios compared with conventional extended Kalman filter–based fusion methods.