Multi-sensor fusion visual SLAM for uncertain observations

With the continued progress of autonomous driving technology, the use of multi-sensor fusion assisted visual simultaneous localization and mapping (VSLAM) has emerged as a new research focus. However, when sensors sensitive to environmental factors output invalid signals, the positioning accuracy of multi-sensor fusion SLAM systems can sharply decline, even leading to system crashes. To address this issue, we propose a multi-sensor fusion VSLAM scheme that can handle uncertain observations. This scheme maintains the stable operation of the SLAM system even when GNSS and image information are lost. By constructing a GNSS simulation neural network based on IMU, we can output position, velocity, and attitude data similar to those measured by dual-antenna GNSS. Subsequently, an error state Kalman filter is used to fuse the network output, GNSS measurement position information, and IMU data. The fusion results assist in the stable operation of the visual SLAM system. Through comparisons with VINS-Fusion on public datasets, experimental results show that our proposed method reduces the average root mean square error by 11.6%. Additionally, simulation experiments under extreme conditions verify the effectiveness of our approach.