Monocular Vision Inertial Odometer Based On Point Line Feature Fusion

Autonomous Navigation Technology of robot in unknown environment has become a key technology.In the research of the theory and method of robot autonomous navigation,a lot of research results have been obtained in the known environment.However,for the unknown environment,although many methods of autonomous navigation have become more and more perfect,there are still many theoretical and technical problems to be solved.Vision based simultaneous localization and mapping(SLAM)is a key technology of robot in recent years.It takes the environment map information and self positioning information provided by robot in unknown environment as the premise of navigation.In dim scenes,using point feature tracking method as the image front-end of visual slam often leads to tracking failure or unable to track;and simply using visual information will make the robot produce increasing cumulative error in longterm movement.In order to improve the navigation and positioning performance of the robot in short-term fast motion and dim environment,this paper takes the mobile robot as the research object,and carries out the integrated navigation research of inertial and visual depth information fusion based on point feature and line feature.The main contents of this paper include:Firstly,aiming at the problem of poor point feature extraction and information redundancy in the scene with rich line features in the dark environment such as corridor,this paper solves the problem by using point features as the main feature and line features as the auxiliary feature.Specifically,Shi Tomasi corner is used to extract point feature information,while LSD(line segment detector)line segment is added to extract line feature,LK(Lucas – Kanade)optical flow method is used to track point feature and LBD(line band descriptor)descriptor method is used to track line feature,so as to ensure the stability of point line data extracted and tracked in dim environment.At the same time,this paper systematically analyzes the causes of the point re projection error and line re projection error in the above process,and more completely deduces the analytical forms of Jacobian matrix and covariance matrix of the camera pose and related feature position,and determines the feasibility of point line through simulation experiments,as well as the operation stability of the system in the above real scene.Secondly,on the basis of vins mono project,the line feature extraction is added,and a monocular visual inertial SLAM algorithm based on point line feature is proposed.The state quantity required at the initialization time of the system is obtained by the visual inertial joint method.Facing the problems of error accumulation and large data dimension in the process of back-end optimization,this paper constructs a new objective optimization function,which integrates the point line re projection error,IMU error and loop detection error,and adopts the edge optimization strategy,which greatly reduces the accumulated error in the operation process of the system and improves the operation efficiency of the system.Finally,aiming at the problem that the result of long-term estimation of the system is not reliable and can not build a globally consistent trajectory and map,the loop detection and relocation research based on bag of words model is carried out.The dbow2 bag of the BREF descriptor is used to detect the loop and optimize the 4-DOF pose map,so as to ensure the consistency of the global trajectory and the map when the system runs in long endurance.In order to verify the reliability and portability of the visual inertia algorithm,we use the camera,PC and car to form a mobile platform,and carry out real scene experiments in the laboratory.The results show that the improved algorithm has higher robustness and better portability.