Abstract This paper focuses on intelligent vehicles and utilizes binocular vision, an inertial measurement unit (IMU), and a global navigation satellite system (GNSS) as sensor data sources to achieve high-precision positioning within a factor graph optimization framework. Based on the classical INS kinematic model, a more accurate IMU pre-integration model that accounts for Earth’s rotation is developed, along with a derived noise propagation error model and a corresponding IMU measurement update process, thereby enhancing pre-integration accuracy. To improve feature tracking and stereo matching, a Shi-Tomasi corner detection method combined with INS-assisted pyramidal optical flow is employed to track feature point motions across frames. Various sensor residual factors are then formulated to construct the factor graph, enabling robust optimal state estimation. The proposed system is validated on public datasets and compared against existing methods under identical conditions, with performance evaluated through absolute and relative trajectory errors. Real-time experiments on actual vehicles further demonstrate the system’s robustness and real-time capabilities, while functional ablation studies verify the advantages of incorporating INS-assisted visual processing and the Earth-rotation-inclusive IMU model. Results confirm the proposed stereo matching approach is highly feasible and effective.
Zhang et al. (2025) studied this question.