Visual-inertial SLAM (Simultaneous Localization and Mapping) is a reliable and effective positioning method in indoor and outdoor environments. However, low-cost sensors and extreme environments (such as low light and weak textures) severely impact the nonlinear optimization within the sliding window, loop closure detection, and graph optimization. To address this, the Pedestrian Dead Reckoning (PDR) algorithm, capable of providing robust pose estimation over a certain time period, is introduced as an auxiliary to enhance overall performance. However, its heading is susceptible to inaccurate readings due to variations in magnetic field anomalies.The paper proposes integrating the velocity, attitude, and stride information provided by PDR into the visual-inertial SLAM optimization, which can improve the localization accuracy by 60%. In extreme scenarios, incorporating the PDR pose information into loop closure detection and graph optimization can improve the localization accuracy by over 30%. Furthermore, utilizing the heading information from visual-inertial SLAM to correct the PDR heading offset caused by magnetic anomalies can improve the localization accuracy by over 20%.
Read full abstract