| In mobile robot positioning technology,relative positioning technology Visual Odometry(VO)is a very important component.From drones,underwater robots,to visual SLAM fields,visual sensors have played an extremely important role..The traditional VO is to perform feature matching through the geometric information of the objects in the picture.By designing various complex mathematical models,under ideal conditions without time constraints,the system state can be accurately estimated across different environments.However,in practical applications,constraints such as sensor errors,system modeling,and complex environments all affect the accuracy and reliability of the mathematical model system.Since the geometric mathematical model addresses the limitations of visual localization algorithms,it is found during the development of deep learning that learning-based VO can be used to solve the defects of traditional mathematical modeling.At the same time,since the inertial sensor IMU and the camera sensor VO have the mutual complementation between fast motion and slow motion,the use of visual inertial sensors to estimate pose has become an important research direction to improve positioning accuracy and system robustness.Aiming at the current development trend of visual-inertial odometry pose estimation,this paper mainly studies the following aspects:The first is to study the mathematical model of VIO.It mainly outlines the coordinate system,rotation theory tools,Lie groups and Lie algebras;studies the geometric model of the camera and its distortion;establishes the IMU motion model;and finally summarizes the public datasets KITTI and EuROC.The second is to mainly study deep learning networks based on VO.Through the analysis of visual positioning requirements,the convolutional neural network was selected and studied,and then the CNN-VO network based on the convolutional neural network was built.and the feasibility of the network to complete the visual positioning was verified.According to the requirements of VO long and short time series,the LSTM variant of the recurrent neural network is added to process the time series information to form a CNN-LSTM-VO network,and the loss function is designed.In order to improve the accuracy of network pose estimation,channel attention and spatial attention mechanisms are added to the network,and a deep network of CNN-LSTM-ATT-VO is constructed on the basis of the original,and training and testing experiments are carried out to verify the performance of the network.The third main research is the fusion algorithm of visual odometry VO and inertial odometry IMU.The overall algorithm framework is proposed,and its defects are found by studying the classic extended kalman filtering algorithm.Therefore,in order to avoid the inconsistency of EKF,the EKF algorithm centered on the robot and completing the calculation in Lie group space is used to complete the pose estimation of VIO.Finally,the experimental verification and analysis.According to the VIO estimation algorithm proposed in this paper,according to the existing public KITTI data set and EuROC data set,the outdoor environment and indoor environment are verified and analyzed by experiments,and the experimental verification and analysis of actual scenes are added to further verify the algorithm.effectiveness. |
