Research On Visual-Inertial Motion Estimation Method Of Planetary Rover Based On Incremental Smoothing

The autonomous navigation and positioning of planetary rovers is the fundamental for their normal operation in an unknown environment.In order to improve the robustness and precision of motion estimation,rovers are generally equipped with a multiple sensors.Monocular vision and inertial sensor are widely used due to their light weight,low cost and complementary advantages.Monocular camera can acquire rich visual information and geometric features while there is inherent scale ambiguity problem and inertial sensor can provide absolute scale information which can be used to rectify the scale drift.This article focus on the visual-inertial motion estimation and study the incremental smooth optimization algorithm.Firstly,from the perspective of probability,a general mathematical model for the visual-inertial motion estimation problem is established.Starting from the maximum a posteriori estimation and given the zero mean Gaussian noise hypothesis and the prior information,the visual-inertial motion estimation is transformed into a maximum likelihood estimation problem,and it is described as a factor graph.The general form of the visual factor node and the IMU factor node are given with the general process model and the measurement model of the visual-inertial motion estimation problem.Then,using the variable elimination method,the factor graph model is transformed into Bayesian network,and further converted to Bayesian tree.The detailed conversion and derivation process are given and the equivalence relationship between variable elimination and QR decomposition was proved.Based on the Bayesian tree,an incremental smooth optimization algorithm is constructed.As an algorithm verification and application,a visual-inertial motion estimation system using artificial features is proposed based on incremental smoothing optimization algorithm.The system uses the April Tag code as artificial feature and the April Tag factor is constructed.The pre-integration technique is used for the IMU measurement.Considering the quality of the IMU,the sample time interval maybe not a constant.Besides,the IMU may not be synchronized with the camera.To this end,a more general pre-integration formulation is given.Finally,extensive evaluations are performed to demonstrate the performance of the visual-inertial motion estimation system,including comparison tests on public datasets,tests on real-world scene datasets and tests on simulated environment based on ROS and gazebo.The evaluation results show that the visual-inertial motion estimation system using artificial features based on incremental smoothing can obtain high estimation accuracy,and it is robust to light and motion blur.Compared to EKF-based algorithms,the proposed system is scalable to large-scale environments and maintain near-constant processing efficiency.The simulation experiment shows the application in planetary rover.