Research On Visual And Inertial Fusion Based Pose Estimation Technique

Pose estimation technology plays an important role in the field of robot automation.And the technologies based on fusion of visual and inertial become popular in the current development of pose estimation due to the complementary nature of the two sensing modalities.In the past few years,most fusion strategies rely on filtering schemes,such as EKF,UKF,IEKF and so on.There are two kinds of fusion categories in pose estimation:loosely-coupled and tightly-coupled.loosely-coupled approaches independently estimate the pose by a vision only algorithm and fuse IMU measurement by Kalman filtering model and then obtain the optimal estimation.Tightly-coupled approaches add the image feature information into the state vector and then obtain the optimal pose of the system through the Kalman filtering model.However,the filter-based methods only linearize once at the current state to solve the Jacobian matrix and cross covariance.So a stronger nonlinear system always has a larger linearization error.And the filter-based methods based on Markov hypothesis are unable to deal with the loop closure problem,resulting to estimation error will grow too large to be eliminated.So the batch non-linear optimization methods become more and more popular in the field of visual-inertial fusion.This paper presents a tightly-coupled pose estimation algorithm based on batch non-linear optimization to fuse vision and inertial data.Our algorithm puts the constraint relationship between IMU visual sensor and landmarks into the cost function and then jointly optimize the cost function in a bounded sliding window iteratively considering all correlations.Visual structure is maintained by keyframes in the sliding window while inertial metric measurements are kept by pre-integration between keyframes.We marginalize the old states in the sliding window to bound the computation complexity.Moreover,we presents a new online camera-IMU extrinsic parameter calibration algorithm to initial the system.At last,We validate the accuracy and robustness of our algorithm on the public dataset and real environment by comparing with the other state-of-the-art algorithms.