A GNSS/SINS/Vision Tightly Coupled Integration With Fault-Tolerant Estimation to Improve Pose Accuracy in Degraded Environments

Abstract

The fusion system of global navigation satellite systems (GNSS), strapdown inertial navigation systems (SINS), and camera has been widely used to provide high-precision positioning service for autonomous vehicles, mobile robots, and other applications. In complex environments such as urban areas, the sensor performance is sensitive to the observation environment, where occlusions and dynamic objects will lead to abnormal measurements. The abnormal measurements in the GNSS/SINS/vision integration system may induce the degradation of estimation accuracy or even divergence. To address the above issue, a GNSS/SINS/vision tightly coupled integration system with fault-tolerant estimation is proposed, where the observation covariance will be adjusted according to the difference between the theoretical and actual variances of innovation and residual. To validate the performance of the proposed fusion system, five in-field experiments are conducted in urban environments. The overall results show that the proposed innovation and residual-based covariance adjustment method can significantly improve the performance of GNSS/SINS/vision tightly coupled integration, whose accuracies are 0.605m, 0.254m, and 0.270m in right, front, and up directions. A detailed analysis is conducted on the degradation environments of GNSS and vision, which further demonstrates the effectiveness of the fault-tolerant estimation based on innovation and residual analysis. Specifically, compared to the results of the state-of-the-art algorithms IC-GVINS and VINS-Fusion, the 3D positioning accuracy of the proposed method has improved by more than 60%.