Research And Application Of Visual Inertial Odometry Based On SIFT Features

As an important technology of autonomous mobile robot and related fields,visual inertial odometry has a wide application prospect.The complementarity of image information and inertial measurement unit(IMU)information can be used to calculate the position and pose changes of sensors more accurately and enhance the robustness of the system.Therefore,the research of visual inertial odometry has very important theoretical significance and application value.Aiming at the problem that the correct matching rate of feature points is low when the visual inertial odometry processes image information,the SIFT(Scale Invariant Feature Transform)feature extraction and matching method is improved.By combining the gradient information of the image to improve the difference between the descriptors in different regions of the actual texture,and using the pixel coordinate information of the feature points,the feature points to be matched between the images collected at adjacent times are screened,so as to improve the correct matching rate of the feature points.Aiming at the problems of poor robustness and scale uncertainty when using a single image sensor to calculate its own pose,a monocular visual inertial odometry is designed by combining with IMU.The Improved SIFT feature extraction and matching algorithm is used to establish the relationship between images at different times.On this basis,the tight coupling method is used to establish the relationship between visual information and pose information.According to the properties of Lie group,Lie algebra and quaternion,the Jacobian matrix is calculated.In order to verify the designed visual inertial odometry,the data set is used for simulation experiment and analysis,and the sensor and embedded processor are used to establish the hardware test platform of visual inertial odometry.The experiments are carried out in indoor and outdoor environment by hand-held mode and robot loading mode respectively.The test results on the public data set EuRoC show that the absolute trajectory error of the designed monocular visual inertial odometry is less than 0.4m on all data sequences,the average time of image processing is 38.86ms,and the average time of pose solving is 35.27ms.In the outdoor circular motion experiment with a total length of 233.720m,the positioning error of the end point is 2.584m,and the ratio of the error to the trajectory is 1.1056%.On the embedded platform,the average pose calculation time is less than 50ms.The simulation and experimental results verify the effectiveness of the monocular visual inertial odometry designed in this thesis.