Research On Robot Localization Technology Based On IMU And Vision Fusion

The precise positioning is one of the prerequisites for mobile robots to perform their tasks.In recent years,positioning technology based on visual information has become an important research topic for autonomous navigation and positioning of mobile robots.However,when the image features are not rich enough or the illumination is not good,the problem of feature loss will occur and further result in poor robustness.For the Inertial Measurement Unit(IMU)sensor,its performance is not affected by light.Even when the visual sensor moves quickly and the image features are missing,IMU can still accurately output the measurement results with high frequency compared to the visual sensor.To a certain extent,it makes up for the shortcomings of the visual sensor.Therefore,visual inertial odometry(VIO)technology based on IMU and visual fusion positioning has received more and more attention.In this thesis,positioning technology of robots based on IMU and vision fusion are mainly concerned.An VIO algorithm based on the improved optical flow method is developed and corresponding performance is verified via public datasets and actual scene data.First of all,the visual positioning system from front-end design to back-end optimization is demonstrated.Referring to the front end of the visual positioning system,key issues such as FAST feature extraction,optical flow tracking without feature descriptor calculation,and PnP pose calculation are analyzed.It indicates that the accuracy of optical flow tracking is an important factor to ensure positioning accuracy.For the back end of the visual positioning system,in order to eliminate accumulated errors,the focus is on loop detection and global pose optimization methods.Secondly,since the captured images in real scenes are difficult to guarantee constant brightness,small motion and other conditions,the commonly used optical flow method may fail to track feature points,induce tracking error,or even completely unable to track when performing pose estimation.To deal with these problems,an improved optical flow method is proposed.Through multiple iterations of feature point tracking and the establishment of image pyramids,the tracking quality of feature points is ensured.CPU+CUDA hybrid acceleration is used to increase the processing speed of the algorithm.In this context,the improved optical flow method can improve both the accuracy of optical flow tracking and the tracking speed.Then,the fusion positioning method of tightly coupled vision and IMU is studied.The IMU pre-integration algorithm is used to process its high-frequency measurement data to avoid repeated IMU integration and costing calculations caused by the change of initial state.With the VINS-Fusion algorithm as the framework,a tightly coupled back-end optimization model is established,and the methods of relocation and global pose graph optimization are studied.After that,the effectiveness of the positioning algorithm is analyzed and verified.Finally,based on the built mobile robot platform,the joint calibration of IMU and binocular camera is conducted,and the fusion positioning experiments are carried out in indoor and outdoor scenes.The experimental results verify the accuracy,real-time and robustness of the positioning algorithm.