A Research Of Autonomous Positioning Algorithm For Mobile Robot

With the development of intelligent robot technology in recent years,self-localization technology has been widely used and researched in the fields of mobile robots,autonomous driving,and unmanned aerial vehicles.Visual odometry is one of the most commonly used methods for autonomous vehicle positioning.This method uses a camera as the main measurement tool,which is simple to deploy,low in cost,and highly adaptable to the environment.It can also work in areas without external signals.The main research contents of this paper are: Lightweight visual odometry based on deep learning and visual-inertial navigation integrated odometry technology based on extended Kalman Filter.Most of the existing visual odometry systems need to be manually calibrated,and in some environments they will lose their original performance or even fail to work properly.The visual odometry based on deep learning requires almost no manual calibration and is more adaptable to the environment,but it will consume a lot of computing resources.Aiming at this problem,the paper studies a lightweight deep learning visual odometry calculation method.The TVNet structure is selected as the optical flow extraction algorithm to shorten the network's calculation time,and the Densenet-like structure is used to improve the accuracy of pose estimation.The paper is tested on the KITTI dataset and the ApolloScape dataset.The results show that compared with similar algorithms,the proposed algorithm reduces the computation time by 50% and improves the accuracy of the rotation estimation on the test data set by 10%.Aiming at the problem that the visual odometry has insufficient estimation accuracy and will fail in a single feature environment,the paper uses the traditional loosely coupled visual-inertial odometry as the framework.Based on the in-depth study of the kinematics of the robot,a loosely coupled fusion algorithm of deep learning visual odometry and inertial measurement unit based on extended Kalman filter is presented.The algorithm of this paper is aimed at the drift of the second integration of the accelerometer,and uses the kinematics model to update the filtering state,which effectively improves the linear modeling error of the robot's short-term motion.The paper is verified on the KITTI dataset.Compared with the traditional Kalman filtering algorithm,the displacement estimation accuracy of the proposed scheme is improved by 17%.