Research On Autonomous Navigation Of Unmanned Vehicles Based On Visual SLAM

The autonomous navigation technology of unmanned vehicles helps the vehicle to complete the driving task independently in the unknown environment and without human intervention.The problem of positioning and tracking is an important prerequisite for the realization of autonomous navigation technology.In recent years,for the positioning of unmanned vehicles,researchers no longer only focus on GPS and high-precision maps,but turn more attention to visual SLAM,which uses image data collected by on-board cameras for positioning.But,because of the complexity and variability of the actual driving environment,especially in low-texture and short-time fast-moving scenarios,the pure visual SLAM autonomous navigation technology can be unable to fulfill the actual positioning accuracy requirements.So as to accomplish accurate positioning of unmanned vehicles in complex scenes,the complementary advantages of IMU and vision and the large number of line features that exist due to insensitivity to illumination changes are combined,this paper takes the visual SLAM autonomous navigation technology as the object,and conducts research on the visualinertial fusion process and line feature algorithm.First,considering the performance and efficiency problems in the visual feature processing stage,the research on feature extraction and matching algorithms is carried out.At this stage,this paper uses ORB and EDLines algorithms to extract point and line features respectively,and then calculate their descriptors BRIEF and LBD and use them for feature matching between frames.Then,the visual constraints between the two frames,namely the reprojection errors of point and line features,are analyzed and constructed,and their Jacobian matrices about pose and landmark are derived.Finally,a nonlinear optimization algorithm is used to estimate the state quantity.Experiments show that,compared with the traditional LSD algorithm,the performance of the EDLines algorithm used in this paper is equivalent but the speed is increased by more than 70%,thereby improving the positioning accuracy and real-time performance of the system.Second,this paper fuses vision and IMU in a tightly coupled manner based on nonlinear optimization.In the fusion process,the IMU pre-integration is used to simplify the calculation.In the initialization stage,this paper proposes a visual-inertial step-by-step joint initialization scheme,which not only considers the uncertainty of the sensor,and uses the maximum a posteriori estimation MAP to transform it into an inertia-only optimal estimation problem,but also solves all inertial Variables at once avoid ignoring correlation due to decoupling,and in pure vision MAP estimation,the monocular vision SLAM initialization scheme is improved,thereby improving the positioning accuracy and robustness of the system.Thirdly,Combining the research on visual-inertial fusion and line features,design the algorithm framework.Based on the ORB＿SLAM framework,this paper introduces IMU measurement data and line feature to design a visual-inertial SLAM autonomous navigation algorithm.In this framework,its state quantities include camera pose,velocity,and IMU bias,etc.,and the reprojection error of point and line features and IMU pre-integration residuals are used to jointly optimize the system state quantities within the sliding window.In pose estimation,different optimization strategies are designed according to whether the map is updated,and the constraint information of the state outside the sliding window is converted into the prior of the state to be optimized in the window through marginalization,thereby enhancing inter-frame constraints and avoiding information loss.In order to correct the accumulated error of the system,loopback detection and global optimization are introduced to ensure the global consistency of the system.Finally,relevant comparative experiments are designed and the performance of the proposed algorithm is evaluated on public datasets.Experiments show that in low-texture and fastmoving scenes in a short time,the algorithm in this paper can effectively complete the tracking and positioning task,with high positioning accuracy and small overall error.Compared with other algorithms,its positioning accuracy and mapping effect are improved.and more robust.