RAC-VIW: A Robust Visual–Inertial–Wheel SLAM With Adaptive Weight Optimization and Feature Point Deviation Compensation

Existing multisensor fusion simultaneous localization and mapping (SLAM) methods typically rely on covariance matrices to model measurement uncertainty. However, anomalous sensor measurements often exhibit non-Gaussian distributions or systematic biases, which are difficult to accurately describe using covariance alone. To address this challenges, this article proposes a multisensor fusion SLAM framework with adaptive weight optimization, which tightly integrates RGB-D images, inertial measurements, and wheel odometer within a factor graph. This system utilizes a weight allocation strategy based on sensor pose increment discrepancies to quantify the degree of anomalies in each sensor. When a sensor anomaly occurs, the estimated pose increments of the anomalous sensor deviate significantly from those of the other sensors. By analyzing these discrepancies, the system effectively detects anomalous sensors and assigns appropriate weights according to the degree of deviation. This approach maintains stable operation even when a sensor fails, significantly improving localization accuracy and robustness. In addition, to mitigate Z-axis trajectory drift caused by feature point deviation during the high-speed rotation of ground robots, we propose a feature point deviation compensation mechanism. Specifically, this method corrects y-axis deviations of feature points in the normalized camera coordinate system, which arise due to optical flow tracking errors, thereby stabilizing optical flow tracking. This approach significantly reduces Z-axis trajectory drift and further enhances the system’s localization accuracy and stability. Experimental results demonstrate that the proposed method exhibits superior performance in complex environments, with the overall system achieving higher localization accuracy compared with state-of-the-art algorithms.