Research On Vision Slam Algorithm Of Mobile Robot Based On The Fusion Of Stero Camera And IMU

In recent years,driverless technology has developed rapidly,both Waymo of Google and Apollo of Baidu use three-dimensional laser sensors to locate and map,but because of its high price,visual slam technology,which is cheap and rich in image information,has become a research hotspot in recent years.This paper focuses on the location and dense mapping of mobile robots,using stereo camera and IMU as sensors.In order to improve the robustness of the camera in the complex environment,this paper uses the scheme based on point and line features to extract point and line features from the image,and uses LK optical flow method and frame to frame matching method to track the extracted point and line features.The depth information of feature is recovered by triangulation principle,and the error function of point line feature projection is deduced by the point line feature common view model.In order to reduce the accumulated error of long-time positioning,the loop optimization algorithm is used to optimize the camera position and pose.Through the test of the data set collected in the complex environment,it is proved that under the premise of real-time tracking,the success rate of image tracking based on the stereo vision odometer based on the point line feature is significantly higher than that of VINS-FUSION.In this paper,a location algorithm based on the fusion of stereo camera and IMU is used to solve the problem that the location error increases when the light is darker and the light changes.During initialization,PNP algorithm is used to solve the initial pose of camera.In order to reduce the influence of IMU bias on positioning,the initial bias of IMU is estimated,and the sliding window algorithm is used to constrain the number of variables to be solved to reduce the calculation.The cost function of tight coupling between stereo camera and IMU is obtained by deriving the preintegration formula and covariance matrix state transfer formula.Through the test of three sequences of data set,the results show that the positioning accuracy of the algorithm of stereo camera and IMU fusion is improved compared with that of VINS-FUSION.When tested in the simulation environment,the absolute trajectory error meets the accuracy requirements of the actual positioning error,which verifies that the positioning algorithm proposed in this paper has a certain practical significance.In view of the problem that the map constructed by location algorithm only contains sparse feature points and dense mapping algorithm based on RGB-D camera can only be used for a fixed indoor scene,this paper uses a dense mapping algorithm based on TSDF fusion to recover the dense depth of stereo image through neural network,and uses TSDF fusion algorithm to solve the problem of dense mapping point cloud redundancy.Through the test of data set,the results show that the precision of dense mapping algorithm based on TSDF fusion is significantly higher than that of Elastic-Fusion.Through the test of data set and simulation environment,it is verified that the dense mapping algorithm proposed in this paper can also do dense mapping for outdoor scene and large-scale environment.