Mobile Robot Trajectory Estimation Based On DNN-Kalman Filtering Algorithm

With the widespread application of unmanned vehicles,trajectory estimation of mobile robots has become a current hot research topic.For different scenes,it is crucial to accurately estimate the true trajectory.In order to accurately estimate the trajectory of mobile robots in different environments,this thesis proposes a method for trajectory estimation of mobile robots based on DNN-Kalman filtering algorithm.The realization of this method can estimate the true trajectory of mobile robots based on IMU(Inertial Measurement Unit)data only.In this thesis,we established an IEKF for Inertial Navigation model,introduced the basic knowledge of Lie group and Lie algebra,Kalman filter and extended Kalman filter,and derived the model formula of invariant extended Kalman filter applied in IMU scene in detail.At present,the noise covariance matrix needs to be set by engineers based on experience.This thesis uses neural network training to directly obtain the measurement noise covariance matrix,and proposes a fusion algorithm based on DNN-Kalman filtering to estimate the trajectory of mobile robots.The key of this method are the Kalman filter and the deep neural network used to output the measurement noise covariance matrix.The state variables,system model and observation model are given,and the formulas of the prediction and update process are derived according to the Kalman filter calculation steps.The measurement value of IMU is selected as the input of the system(deep neural network and Kalman filter),and the system output is the update state.Simulation experiments were carried out on the algorithm,and the accuracy of system trajectory estimation was verified with IMU data from KITTI data set and KAIST data set respectively.The results showed that the KITTI dataset was validated better and with less error than the KAIST dataset.Finally,based on the experimental platform of mobile robot vehicle,the data are collected and verified in the actual environment,and the better trajectory estimation effect is obtained.