Abstract: In low-altitude complex environments, GNSS observation quality is strongly non-stationary, and abrupt epoch-to-epoch changes can trigger pronounced solution jitter and accuracy degradation in loose-coupled strapdown inertial navigation system (SINS)/GNSS integration with low-cost sensors. To enhance robustness and positioning accuracy under such abrupt quality variations, an exponentially smoothed sequential soft chi-square adaptive measurement-weighting method is proposed. Within an error-state Kalman filter (ESKF) framework, component-wise normalized innovation statistics are constructed to perform consistency checking, and a dual-threshold soft regulation mechanism is employed to continuously down-weight abnormal measurement components, thereby avoiding information loss caused by hard rejection. Meanwhile, an exponentially weighted moving average is applied to the online-updated measurement covariance matrix, which suppresses high-frequency switching of measurement weights and filter gains in abnormal intervals and consequently reduces oscillatory behavior of the fused solution. Experiments conducted on an open-source MPU6500-GNSS dataset with pronounced GNSS fluctuations demonstrate that the proposed method significantly mitigates horizontal error spikes and improves trajectory smoothness. Over the 1100-1800 s interval, the root mean square (RMS) of the east-position error decreases from 19.303 m to 7.986 m and the maximum decreases from 99.512-55.325 m; the RMS of the 3D position-error norm decreases from 19.411-10.201 m and the maximum decreases from 99.523-63.223 m. The results indicate that the proposed scheme effectively limits the adverse impact of abrupt GNSS quality degradation, offering a simple, computationally light, and statistically interpretable strategy for stable and accurate low-cost inertial/GNSS positioning.