Research On Autonomous Localization Method Of Mobile Robot Based On Visual Odometry

Simultaneous Localization and Mapping(SLAM)of mobile robots has become one of the hotpots in the field of computer vision.As the most important part of computer vision SALM,the visual odometry plays a key role in the localization and navigation of mobile robots,and is also the most widely used localization method for mobile robots.Visual odometry acquires image frames through calibrated cameras,detects and extracts image frame information for matching optimization,and then uses computer vision geometry and camera parameter model to estimates the 6-DOF pose of mobile robots.In this paper,the basic theory of algorithm of each link of binocular visual odometry is studied systematically,the related algorithm of visual odometry is improved,and a realization method of binocular visual odometry is proposed,which overcomes the problem of low precision and high error of traditional visual odometry.Its basic theory mainly includes the image features extraction,image features matching,optimization and pose estimation of mobile robots.Firstly,the tracking characteristics of feature optical flow are studied based on the variable threshold ignition characteristics of neurons of Pulse-coupled Neural Networks(PCNN).The advantage of the variable threshold ignition feature is that it can self-adaptively mark the edge features of the input small-scale continuous image frames,and the marked edge features have rotation invariance,intensity invariance,scale invariance and distortion invariance,so that the tracking feature points also have these characteristics.The algorithm is verified by the Middlebury datasets,KITTI datasets and real datasets.The experimental results show that the features of variable threshold ignition can be used to calculate more dense optical flow field,which indirectly proves that optical flow has tracking characteristics.Then,based on the tracking characteristics of LK feature optical flow and Oriented FAST and Rotated BRIEF(ORB)algorithm,a fusion matching algorithm LK-ORB is proposed.The LK-ORB algorithm first uses pyramid feature optical flow to track ORB feature points in the local feature window and calculates the displacement vectors of the feature points.Then a coarse-to-fine matching strategy is proposed to match the tracked feature points.The algorithm is validated by the Málaga datasets and New College datasets.The experimental results show that the feature matching rate of the fusion matching algorithm can reach 98%,and the coarse-to-finematching strategy has more precise result.Finally,a robust and efficient pose estimation algorithm for mobile robots is proposed to solve the problem of low accuracy and high error of feature matching in traditional algorithms,which leads to the decline of positioning accuracy of mobile robots.The algorithm constructs binocular visual odometry system based on initial pose estimation thread and optimization thread respectively,in order to minimize the re-projection error after feature tracking,a closed-loop matching strategy is introduced in the initial pose estimation thread,that is to say,the feature points in two pairs of continuous binocular images are cyclically matched.The pose estimation is completed by calculating the re-projection error of matching points and BA optimization.The KITTI datasets and EuRoC datasets are used to verify the performance of the algorithm.The experimental results show that the proposed algorithm is more challenging in short-time motion pose estimation,with the average translation error of 0.98% and the rotation error of 0.0028 deg/m.For further confirming the effectiveness of the algorithm,this paper uses the self-built unmanned vehicle platform to collect outdoor scene images and construct datasets for experimental testing and comparing with the vehicle GPS positioning results.The experimental results show that the proposed algorithm can estimate the pose which is basically consistent with the real pose.