Pure IMU localization for intelligent platforms with CNN adaptive invariant extended Kalman filter noise fusion

In the context of intelligent vehicles, low- and medium- precision Inertial Measurement Units (IMU) are plagued by high levels of noise and considerable output uncertainly. When positioning and attitude estimation rely solely on IMU data, errors rapidly accumulate over time. To address this issue, this paper introduces a Convolutional Neural Network (CNN)-based noise-adaptive invariant extend Kalman filter (IEKF) vehicle localization. The proposed approach develops CNN models tailored for IMU measurement data as well as the process noise and observation noise in the IEKF. An enhanced CNN architecture and convolution mechanism are designed to dynamically adjust the covariance matrices associated with both process noise and observation noise in response to varying IMU input. This integration with IEKF principles ensures real-time positioning while achieving high accuracy in position prediction. The proposed method was tested and validated on 16 IMU sequences from the KITTI dataset, resulting in a relative translation error performance improvement ranging from a minimum of 10 % to a maximum of 24 % when compared to four existing methods. Additionally, its performance was further evaluated through various metrics including cumulative distribution of errors, root mean square error, and absolute position error. Trajectory tracking experiments further demonstrated that the proposed method produces smoother localization curves and more stable positioning performance.