Monocular Visual-Inertial System Based on Adaptive Scale Recovery of Structured Features

With the development of technology, the automation industry has gradually improved, and the application of the visual-inertial system has become increasingly prosperous. Self-driving cars and autonomous mobile robots often use the visual-inertial system as a state estimator, and their image information can be used in the artificial intelligence part to provide more information about the robot. The combination of a monocular camera and an inertial sensor has the characteristics of light weight and low power. Monocular Visual-Inertial SLAM is a system that relies on monocular cameras and inertial sensors to perform localization and mapping. Monocular Visual-Inertial SLAM restores its vector to scale with the help of known extrinsic parameter information. However, compared with other methods, it is necessary to be careful in estimating the gyroscope bias of the inertial sensor. The bias and noise of the inertial sensor will directly affect the results of pre-integration, as well as the accuracy of the visual-inertial scale recovery. This phenomenon is called scale uncertainty, which results in feature points with uncertain scale or depth, rapidly reducing the accuracy of pose estimation. This paper proposes an architecture of visual-inertial based on Structure-Scale Adaptivity. In this study, the tightly coupled residual between the structure scale and the pose is designed to provide adaptivity according to the feature conditions, providing system structure and scale constraints, and thus suppressing the problem of scale uncertainty. This study is compared with state-of-the-art algorithms and succeeds in scenarios where other algorithms have failed. It also obtains the best accuracy results in the overall test. Finally, we evaluate our proposed method on VCU-RVI and KITTI datasets, demonstrating the proposed algorithm's effectiveness and feasibility for monocular visual-inertial SLAM.