Research On Monocular Vision Pose Estimation Method Based On Point And Line Feature Fusion

As one of the core technologies in the field of computer vision research,monocular visual odometry(Visual Odometry,VO)can estimate the surrounding environment of the monocular camera and its own posture only through the image data collected by the monocular camera.It provides important technical support for carriers such as drones to complete navigation and positioning tasks in an unknown environment.The traditional monocular visual odometer based on point features is greatly affected by the environment.In the environment with low texture and large changes in lighting,there will be problems of reduced positioning accuracy and even tracking failure.In response to the above problems,this paper introduces line features in the frame of monocular visual odometer based on point features and improves the feature matching algorithm to reduce the impact of environmental changes on the accuracy of visual monocular visual odometer.The main research work of this article is as follows:(1)A method for eliminating mismatching point features between frames based on grid evaluation is studied.In the point feature processing process based on the point line feature monocular visual odometer,the pixel points in the image coordinates have a spatially similar motion smoothness constraint to evaluate the similarity of the point feature matching clusters in the grid area And eliminate matching clusters with poor similarity.Compared with the traditional method of using the nearest neighbor algorithm for false matching rejection,the gridbased false matching rejection method retains more effective matching points,and the calculation time is only half of the former.(2)A line feature matching method based on geometric constraints is proposed.In the process of line feature processing based on the point-line feature monocular visual odometer,a line feature filtering process based on a pixel gradient filter is introduced,and geometric constraints such as the angle between the line features and the projection are used to implement a new frame Interline feature matching method.The inter-frame line feature matching experiment shows that,compared with the traditional descriptor-based inter-frame line feature matching method,although the number of inter-frame line features matched by the proposed method has decreased,its internal rate has increased by 10% over the former,The matching time was shortened by 22.4%.(3)A pose estimation method of fusion point-line feature map optimization model is studied.In the traditional monocular visual odometer based on point features,this paper adds the constraint of reprojection error from the space line to the image frame on the existing graph optimization model,and carefully deduces the constraint in Euler space for solving the optimal Correlative Jacobian matrix for state estimation.It can be seen from the pose estimation experiment in the TUM dataset that the position estimation method using the graph optimization model proposed in this paper improves the positioning accuracy by 13.1% compared to the pose estimation method using the original image optimization model,and the positioning error after introducing the scale true value Within ± 2cm.(4)Integrate the above algorithm to achieve a complete monocular visual odometer with fusion points and line features,with features such as feature tracking,inter-frame feature matching,mapping and local pose optimization.Compared with the traditional monocular visual odometer based on point features,the positioning accuracy of the odometer proposed in this paper is improved by 10% in the original Eu Ro C data set,and after processing,the Eu Ro C data set that simulates the environment of lighting changes is increased by 15%.After the value,the positioning error is within ± 10 cm.In the measured data experiment,the loop error of the monocular visual odometer proposed in this paper is less than 2%,which is better than the traditional monocular visual odometer.