CNN-based Invariant Extended Kalman Filter for projectile trajectory estimation using IMU only
Alicia Roux, Sébastien Changey, Jonathan Weber, Jean-Philippe Lauffenburger · 2021
This paper presents a new method to estimate the position, velocity and orientation of a projectile using only accelerometers, gyrometers and magnetometers. This algorithm is composed of an Imperfect Right-Invariant Extended Kalman Filter (R-IEKF) and a Convolutional Neural Network (CNN) to dynamically optimize the measurement noise covariance matrix. This time-varying covariance matrix is always adjusted to each phase of the projectile flight. To evaluate the algorithm performances, mortar fire simulations are performed and a comparison between the estimated trajectory and the reference is shown. The estimates obtained with an Imperfect R-IEKF outperform Dead Reckoning estimates. In addition, the joint use of an Imperfect R-IEKF and a CNN enables to significantly reduce the estimation errors of the projectile trajectory.