Research On Visual Inertial SLAM Algorithm Based On Point-line Feature Fusion

Simultaneous Localization and Mapping(SLAM)technology is one of the essential technologies in the study of mobile robotics and autonomous driving technologies,which have become a hot topic of research today.The vision-based SLAM technology is more of a research challenge in this field.Most of the current visual SLAMs use point features as the image information of the visual front-end for localization and tracking.However,in the scenes of weak texture and fast motion,the SLAM system has a large localization error in such scenes due to the failure to extract enough point features,and even the running trajectory is lost.To address such problems,this thesis proposes a visual-inertial SLAM system(Visual-inertial System,VINS)based on point-line feature fusion,which adds line feature constraints to the visual front-end and couples with point features to achieve high-precision localization of the system in complex scenes.The main research contents and methods of this thesis are as follows.First,in the front-end feature processing stage,for point features,Shi-Tomasi angle points are extracted and then tracked using the improved LKT(Kanade-LucasTomasi)optical flow;the number of matching point features is improved.For line features,a short line suppression strategy is set on the basis of the original line segment extraction LSD(Line Segment Detector)algorithm,which improves the quality and efficiency of line feature extraction.And the LBD(Line Band Descriptor)descriptor matching algorithm is improved,and a sampling point-based line matching method is proposed,which combines two matching methods of frame-to-frame and frame-toplane and is used to improve the accuracy of line matching.The Plücker coordinates and orthogonal representation is a method to describe spatial line features,and this method can be used to construct a reprojection error model for line features.Also,the algorithm uses an inertial measurement unit(Inertial Measurement Unit,IMU)to provide position and orientation aids.the pre-integrated measurements of the IMU can be used to constrain the relative poses to construct the IMU residuals,which are passed to the back end for nonlinear optimization.Then,visual inertial initialization of the system is performed,and detailed derivations of the state quantities to be initialized are performed,and these state quantities are passed to the back-end for optimization as initial values.In the back-end positional optimization,a sliding window model is used to integrate a priori information,loopback detection residuals,point-line feature reprojection errors and IMU residuals for each partial residual to construct the objective optimization function and solve it by global BA(Bundle Adjustment).Finally,a complete VINS system with point-line feature fusion is built,and the algorithm of this system is compared with the classical algorithms VINS-Mono and PLVIO on the public dataset EuRoC for various aspects of experiments and performance analysis,including the operating trajectory and absolute positional error of each algorithm.According to the description of the experimental results,it is shown that the visual inertial SLAM system with point-line feature fusion proposed in this thesis can have strong robustness in complex indoor and outdoor scenes.The algorithm is not only able to extract and match more point-line features,but also to achieve more accurate global positional estimation.To demonstrate this,the researchers integrated the algorithm into the INDEMIND camera and experimentally validated it in dim,weakly textured and empty scenes with complex structures.The experimental results show that the algorithm exhibits strong robustness in these complex scenes,thus demonstrating the effectiveness and usefulness of the algorithm.