<p>To improve the accuracy and stability of bridge weigh-in-motion (BWIM) systems under complex traffic conditions, this study proposes a Kalman filter-based method with adaptive observation noise modeling. Unlike conventional approaches that rely on empirically tuned noise parameters, the proposed method employs variational modal decomposition to extract the quasi-static component from bridge responses, which is then used to estimate the observation noise covariance matrix in a data-driven manner. This strategy enhances the reliability of Kalman filtering and reduces sensitivity to uncertain measurement noise. Extensive numerical simulations were conducted to evaluate the effects of noise intensity, vehicle speed, sensor configuration, and noise modeling strategies on estimation accuracy. The proposed method demonstrates high accuracy and stability under various traffic scenarios, outperforming regularization-based techniques in both numerical and experimental validations. Laboratory experiments on a scaled bridge model and field tests on a full-scale bridge further validate the effectiveness of the method, demonstrating strong robustness against noise, road roughness, and complex multi-vehicle interactions. The results suggest that the proposed methodology provides a practical and scalable solution for real-world BWIM applications, with improved adaptability and minimal reliance on empirical parameter selection.</p>