Research On Visual Inertia SLAM Algorithm Using Point And Line Features

Simultaneous Localization and Mapping(SLAM)has a wide range of applications in virtual reality,augmented reality,and other fields.The point feature-based SLAM system is difficult to extract enough feature points,or the feature points are unevenly distributed in scenes with low-textured,severe lighting changes,and fast motion,which leads to the failure of location.However,the line features have a good performance in the above scenes.This paper focuses on the visual-inertial SLAM algorithm combining point and line features.First,to solve the problem that cannot extract enough feature points in textureless or low-textured environments,which leads to the reduction of positioning accuracy or even failure,a visual-inertial SLAM system using point and line feature is proposed.In the back end,a cost function is constructed by the reprojection error of point features,the reprojection error of line features,the prior residuals introduced by marginalization,and the IMU measurement error.The pose of the robot and the coordinate of the features are obtained using nonlinear optimization of the cost function.Experiments show that the visual-inertial odometry using point and line features can get an environment map with point and line features.Secondly,in view of the problem that the cumulative error cannot be eliminated in the visual inertial odometry,a loop detection method using the point and line feature based on the bag of words is introduced.The dictionary of point features and line features are trained separately.In the process of loop detection,the results of the two bag-of-words models are weighted to obtain the final loop candidate frame.If there is a similarity score greater than the threshold,the corresponding keyframe is a loop candidate frame.The loop candidate frame is added to the sliding window for fast relocation after geometric verification.When the frame marginalized out of the sliding window has a loop relationship with the frame in the keyframe database,all the frames in the keyframe database are globally optimized.In this way we can get a globally consistent pose estimation.Experimental results prove that the loop detection based on point and line features can effectively eliminate the accumulated error of the odometry,and the effect is better than that of only using point features for loop detection.Finally,in view of the impact of dictionary size on the results of loop closure detection,a loop closure detection algorithm using binary search tree is proposed.Randomly select a bit index in the binary descriptor as the node and assign the descriptor to the left node or the right node according to the corresponding value of the bit of the node.Repeat the above steps.Each bit index can only appear once,and finally the descriptors are placed in the leaf node.The matching of the descriptors uses the Hamming distance.If the Hamming distance between the descriptors is smaller than the threshold,it is considered that the features corresponding to the two descriptors are the same,and finally the loop candidate frame is selected by the voting scheme.Through experimental verification,the proposed loop closure detection algorithm is more efficient than the detection method of the bag-of-words model under the premise of ensuring accuracy.