Research And Design Of Odometer System Based On Monocular Vision And IMU

With the development of computer vision and micro inertial navigation technology,the navigation system based on the fusion of vision and micro inertial navigation has been widely used in smart mobile robots,autonomous navigation vehicles,and unmanned aerial vehicles.Both inertial sensors(IMU)and vision sensors have the advantages of small size,light weight,low cost,etc.At the same time,inertial sensors(IMU)and vision sensors have obvious complementarities.The complementarity between them is mainly reflected in:IMU can provide short-time and high-precision attitude information,which provides a good solution for fast movement,and data of monocular vision will not drift in the slow movement to solve IMU drift problem.This thesis studies an autonomous navigation system based on vision and IMU fusion which is usually called Visual Inertial Odometry(VIO).Combining the features of monocular vision and IMU,an unscented Kalman particle filter(UPF)framework is used to design a monocular visual inertial odometry system based on close-coupled mode.This thesis firstly deduces the attitude,velocity,position updating equations and error equations of the Strapdown Inertial Navigation System,and focuses on the simulation comparison of the attitude calculation algorithms based on the three-subsampled compensation method and Runge-Kutta method,and the advantages of the Runge-Kutta method are verified.Secondly this thesis studies the key technologies based on monocular camera vision system positioning,including the linear model and non-linear model of the monocular camera,the calibration of the monocular camera,the polar geometry constraint in the computer vision and the trifocal tensor technique,image feature extraction and adjacent frame image feature tracking algorithm.Feature extraction and tracking algorithms are improved via using FAST feature point extraction and LK optical flow tracking method,non-maxima suppression algorithm(NMS)and a random sample consensus algorithm(RANSAC).Tests are operated using the image date practically acquired with the Logitech monocular camera to verify the improved image feature extraction and image feature tracking algorithm.Furtherly,the filtering algorithms of monocular vision and micro inertial navigation are studied.According to the relationship between the IMU kinematics model and the multi-view geometry constraint relationship,the state equations and observation equations of the monocular visual inertial odometer system are established.An inertial visual odometer system based on the RUPF filter algorithm is designed based on the random sampling consensus algorithm and the unscented Kalman particle filter algorithm.Finally,the fusion algorithm is verified using the data collected by the KITTI dataset platform and the mobile robot platform based on Logitech/MPU6050.