Robust Visual-Inertial Odometry Based on Deep Learning and Extended Kalman Filter

Vision-inertial odometry navigation system is a low-cost, lightweight, continuous and reliable navigation and positioning method. In order to obtain the accurate and reliable navigation information, the navigation system has to confront the challenge of environmental interference. Due to the unavoidable challenges of turning, accelerating ego-motion and nontextured, dynamic scene for image processing, there is random interference caused by ego-motion uncertainty, which makes the estimation algorithm divergent and positioning unreliable. The purpose of this paper is to develop a robust vision aided inertial navigation strategy, which can be divided into front end and back end. The front end uses a visual deep learning framework based on recurrent neural network for end-to-end state estimation. The back end applies the extended Kalman filter in vehicle coordinate system, and combines the degree of abnormity measuring the uncertainty of the system online in order to dynamically adjust the filtering method. The experiments using KITTI dataset on the unmanned ground vehicle were tested under the drastic change of vehicle movement state and environment. The results showed that the robust vision-inertial odometry navigation system has robustness and adaptability to resist external interference, and can improve the positioning accuracy of unmanned ground vehicle.