Research On Key Technologies Of High Precision SLAM With Multi-Sensor Fusion

As an important part of the current mainstream self-driving technology,robotic perception is based on simultaneous localization and mapping(SLAM).Monocular SLAM is one of the research hotspots and widely used.However,there is a problem of scale uncertainty in monocular SLAM,the monocular odometry DSO has uncertainty of the scale,and the accumulated error in pose estimated will cause the inconsistency of pose scale information,so that the high-precision pose estimation cannot be performed.Aiming at this problem,this paper proposes a visual inertial odometry combining with DSO and IMU sensor,which completes the high-precision pose estimation from the front-end and the back-end.In the initial pose estimation of the front-end,a de-distortion inter-frame pose estimation algorithm is proposed.The distortion parameters are added in the inter-frame matching process and the photometric error function is constructed to solve the visual initial pose.In the back-end pose optimization,the construction rules of global state variables are proposed,and multiparameter optimization criteria are established to reduce the error accumulation of each variable.A multi-parameter sliding window pose optimization algorithm based on tight coupling is proposed.The scale factor is designed to maintain scale consistency between visual and inertial pose information.The priori corresponding to the specific frame is selected to construct an overall objective function for a prior,visual and inertial information.Through the above research,estimate the objective function and get the updated value of the global state variable.Iteratively updates the global state variables to reduces the error accumulation,and achieves the high precision of the system pose calculation and the robustness of the motion process.The main work of this paper is as follows:1.Aiming at the problem that the position of the pixel point projection in image matching is deviated because of distortion.In the initial estimation of the pose,the de-distortion interframe pose estimation algorithm is proposed.The distortion parameter is added to the image matching to construct a visual photometric error function to solve the visual initial pose;the sample values of the IMU module are pre-integrated.The addition of distortion parameters makes the pixel point projection position more accurate,the matching is better,the accuracy of the optimization of the pose is improved,and the pre-integration of the inertial information avoids the repetitive integration phenomenon caused by the frequency frame optimization state frequency change caused by the high sampling rate of the IMU.2.Aiming at the problem that the accuracy of each parameter in the photometric error and the IMU pre-integration of the front-end is affect the accuracy of the back-end pose optimization,the global state variable construction rules are proposed.The selection principle is as follows: IMU pre-integral item,IMU bias;geometric element covered by photometric error,distortion parameter;scale factor,inverse depth of map point.The global state variables can clarify the information of each variable that needs to be solved in the back-end pose optimization,reduce the error accumulation of each variable and improve the accuracy of pose.Compared with the translation drift and scale drift of the unfused IMU system,the experimental data has a smaller drift and the accuracy is further improved.It proves that the optimization of each variable is more accurate and can provides a better initial pose for next frame optimization.3.In the jointly optimizing the pose of the vision and inertial data in back-end,the scale uncertainty of the monocular leads to the accumulation of error in the pose optimization,the inconsistency of the front and rear pose scale information.A multi-parameter sliding window pose optimization algorithm based on tight coupling is proposed.Firstly,according to the initial pose of the front end,the keyframes are selected,the inter-frame photometric error item is constructed,and the IMU sample values in the keyframe corresponding time are preintegrated to construct the inertial residual item.Secondly,the scale factor is designed to maintain scale consistency between the pose of the photometric error and inertia residual;finally,the priori corresponding to the specific frame is selected to construct an overall objective function for a prior,photometric error and inertia residual;Estimate the objective function and get the updated value of the global state variable.Iteratively updates the global state variables to obtain the high-precision pose at the current moment.The experimental results show that the proposed algorithm can obtain high-precision pose results by comparing the experimental data of the pre-fused IMU system with APE and RPE.Through the above work,the visual inertial odometry of the DSO and IMU sensor proposed in this paper is demonstrated.The comparison between the experimental data of the visual inertial odometry and the unfused IMU system and the current mainstream visual inertia method shows that the proposed method has high-precision pose estimation results after the fusion of IMU.