IMU auxiliary tracking model for monocular vision inertial positioning

Abstract

The invention discloses an IMU auxiliary tracking model for monocular vision inertial positioning. The IMU auxiliary tracking model comprises an initialization stage and a stage after initialization is completed, wherein during the initialization stage, pre-integration is carried out on original data acquired by an IMU on the basis of setting initial translation of a current frame at a constant speed model, an IMU priori rotation is obtained by calculation, and an initial pose of the current frame is acquired; and after initialization is completed, prior speed information of the pre-integration of the IMU is added on the basis of providing the initial pose, a tracking model is established according to calculated initial states, and accurate positioning is achieved. According to the IMU auxiliary tracking model, the influence of strenuous exercise and the like can be effectively overcome.

Description

technical field [0001] The invention relates to the technical field of visual positioning, in particular to an IMU-assisted tracking model for monocular visual inertial positioning. Background technique [0002] Monocular vision-inertial tightly coupled positioning technology has developed rapidly in recent years, from the initial tightly coupled method based on Kalman filter to the method based on BA (Bundle Adjustment) optimization and graph optimization. [0003] Anastasios I. Mourikis proposed a typical extended Kalman filter-based tightly coupled model MSCKF in 2007, which has since opened the visual-inertial tightly coupled positioning method. Same as the traditional SLAM framework, the tracking model of MSCKF is a basic visual reference frame model. In the case of tracking, the prior data of the IMU is not used, so the system is very easy to fail to converge, resulting in poor positioning accuracy. Raul Mur-Artal, the author of foreign ORB_SLAM2[10], proposed a visua...

AI Technical Summary

Problems solved by technology

Same as the traditional SLAM framework, the tracking model of MSCKF is the basic visual reference frame model, and the prior data of the IMU is not used in the tracking situation, so the system is very easy to fail to converge, resulting in poor positioning accuracy

Raul Mur-Artal, the author of foreign ORB_SLAM2[10], proposed a visual-inertial monocular SLAM technology with map reuse in 2017. This algorithm is based on the open-source ORB_SLAM2 framework in 2015, and proposes a robust visual-inertial local algorithm based on ORB_SLAM. The BA optimization model also proposes a simple visual-inertial joint initialization method, but this method requires about 15 seconds of initialization time, so a stable visual front-end foundation is required before the initialization is successful to ensure the stability and accuracy of the initialization. However, ORB_SLAM uses a simple uniform tracking model during initialization, which cannot be initialized correctly when the camera moves violently, and even tracking fails.

The Hong Kong University of Science and Technology Robotics Institute proposed VINS-Mono in 2017: a robust monocular visual inertial state estimator, VINS (Visual-inertialsystem) is based on a six-degree-of-freedom state estimation of a monocular fisheye camera and a low-cost IMU sensor In terms of vision, the sparse optical flow method is used for tracking and positioning. In the paper, a robust visual-inertial joint initialization process and recovery process are proposed, as well as a tightly coupled nonlinear optimization method based on sliding windows. However, the entire system uses It is the most basic reference frame tracking model, which cannot overcome complex motion environments

Images