Research On Localization Algorithm Based On Fusion Of Stereo Camera And Inertial Sensor

In recent years,the demand for intelligent robots in various fields has increased rapidly.In some cases,the environment is relatively complex.Under these complex conditions,it is difficult to receive these robots’ GPS signal in real time.This has made it challenging in obtaining the pose information of the robots and the navigation as well as control of the robots.Therefore,more and more researchers are beginning to focus on the localization scheme based on visual SLAM(Simultaneous Localization And Mapping).However,the output frequency of visual SLAM is low,and the robustness is poor in fast motion or rotation,how to improve the robustness of the algorithm is an important task of visual SLAM.The main research contents of this paper are as follows:(1)Aiming at the disadvantage of poor robustness of visual SLAM,stereo VIO(Visual-Inertial Odometry)is presented,which combines the stereo cameras and inertial measurement unit(IMU)of low cost.The output frequency of IMU is high,and the calculation of rotational motion in a short time is relatively accurate,the problem of low output frequency of the visual SLAM can be solved,and the accuracy and robustness of the pose estimation can be improved.(2)The calculation of visual SLAM algorithm is large,it is hard to use high-quality sensors and powerful processors in some cases due to size and cost limitations.In this work,to decrease computation cost,we use the FAST feature detector to improve efficiency and the KLT sparse optical flow algorithm to track features.Additionally,in order to reduce the mismatch caused by optical flow tracking,this paper reduces the amount of mismatch through the constraint between stereo camera,and we perform circular matching between the previous and current stereo image pairs to remove outliers in the stereo matching and feature tracking steps,thus reducing the mismatch of feature points.In the objective function of the back-end optimization,the stereo camera constraint information is added,thereby improving the robustness and accuracy of the system.(3)Finally,to evaluate the proposed algorithm,this work compares monocular VIO based on optimization and stereo VIO based on filtering using the public EuRoC dataset.Experimental results demonstrate that our method has higher localization accuracy.From the experiment on the hardware platform,it can be seen intuitively that the system performs well and has better robustness in continuous long-distance localization after loop detection is turns on.In this paper,a map construction experiment is performed based on the localization algorithm.The camera pose obtained by the visual-inertial algorithm is merged with the stereo camera depth map to obtain a dense point cloud map in 3D space,and the point cloud map is converted into the actual scene octree through a third-party library.The constructed map results can also reflect the accuracy of pose estimation derived from the proposed algorithm.