Monocular Simultaneous Localization And Mapping Towards Dynamic Environments

In applications such as robotics,Augmented reality/virtual reality and other agents,SLAM is an important technology to support the realization of above tasks.Monocular cameras provide the most abundant information for SLAM and widely used due to their small size,light weight and low energy consumption.However,due to the limitation of numerical model calculation,most monocular SLAMs assume that static scenes take dominant in environments,which greatly limits the application of visual SLAM in real scenes and make it unable to meet the ever-developing needs of navigation,cruise and detection,etc.At the same time,The basic physical information of the objects in the scene,such as position,speed and rotation are the basis for the planning and decisionmaking of the sensor-body,This is also the necessary links required for nomal work in areas such as robot,planetary exploration,which triggers prominent requirements for monocular SLAM to percept and track moving objects within scenes..This dissertation focuses on dynamic scenes and conducts in-depth research on the robustness of monocular SLAM and multi-target tracking based on monocular camera.The main research content of this paper includes the following aspects:(1)Research on robustness of monocular SLAM in dynamic environments.Such problem regard moving objects as elements that interfere with the operation of SLAM system and eliminated moving objects in the process of positioning and mapping.The existing SLAM deal with this problem by assuming that the static model has the largest number in interior points,then solve camera motion model using the random sampling consensus algorithm.In dynamic environments,the calculation time of this method increases exponentially as the propotion of interior points of static model decreases,and the algorithm fails when the proportion of interior points is lower than 50%.In response to this problem,this paper proposes a dynamic monocular SLAM system(Monocular SLAM towards dynamic environments,DM-SLAM).DM-SLAM uses static feature extraction-maintenance mechanism for motion scenes to ensure the robustness of monocular SLAM system.In the initialization thread of static feature extraction,DM-SLAM proposes a sampling consistency algorithm(Distribution and local-based RANSAC,DLRSAC)based on neighborhood model calculation to extract static features in the scene.DM-SLAM verifies the effectiveness of the algorithm in our data and in the Technical University of Munich(TUM)data and copare with the RANSAC,ARSAC algorithm to verify the superiority of the algorithm in the accuracy and completeness of static feature extraction.In the static feature maintenancestage,the main contradiction of the system is that if we guaranteed the map is static,the newly added map points need to be geometrically tested in more frames,which will redue the speed of system tracking and the accuracy of pose resolution.In order to balance this contradiction,DM-SLAM proposes a neighboring mutually exclusive alternate map point selection strategy.Finally,this paper makes comprehensive comparison between DM-SLAM and the classic SLAM system ORB-SLAM in the TUM data set,including the evaluation of relative rotation error,relative trajectory error and absolute trajectory error.In the above evaluation,DM-SLAM significantly improved the performance of ORB-SLAM.Among them,in the motion scene video sequence fr3?walking,the square root error and standard deviation of the absolute trajectory error have increased up to 96.42% and 98.45%.(2)Research on multi-target tracking of monocular SLAM.Such problem regards moving objects as task object and need to recover the position and geometric information of the moving objects.The multi-target problem has the characteristics of“chicken-egg” cycl,and the pose recovery based on monocular vision is inherently unsovable.It is usually assumed that the moving target and the camera(or the subject carrying the monocular camera)are coplanar to address the unsovable problem,but this constraint is difficult to meet in practical applications,resulting in low accuracy of monocular SLAM multi-target pose recovery.Aiming at the problem of monocular SLAM multi-target tracking,this paper proposes a visual inertial plus semantic monocular SLAM(VIPS-Mono)system to solve the camera motion scale,targets motion scale solution,and multi-motion target segmentation problem.Aiming at the limitation of the coplanarity hypothesis,this paper proposed a vehicle depth and height joint estimation algorithm,which makes full use of the vehicle height observation consistency constraint,alliviating the coplanarity assumption,and finally combines the feature matching pair information belonging to the moving object to achieve high precision and reliability six-degree-of-freedom target motion solition.Finally,this paper conducts sufficient experiments on VIPS-Mono in the autonomous driving data set KITTI,including the comparison of the robustness of the initialization process and the resolution accuracy of the camera motion scale with the VINS-Mono system.A detailed comparison between direct estimation and joint estimation of target vehicle depth and height was made in the calculation.The depth estimation error of the joint estimation algorithm does not exceed 0.8 m on average and the average height error does not exceed 0.16 m,which meets the application requirements of unmanned driving in experiment scenes.Finally,the target vehicle trajectory of VIPS-Mono is compared with the true value in detail.Compared with the true value,VIPS-Mono's estimation result of the target vehicle trajectory is smoother,which is in line with the real movement of the target vehicle.