A paper introduces DFIO, a method for tracking movement from inertial sensors, which drift over time as sensor noise builds up. DFIO builds the update step of an error-state Kalman filter into the training of the neural network, so the two are tuned together.
On a public dataset of people walking, it cut mean absolute translation error by 13.2% and one-second relative translation error by 39.6% against the TLIO baseline. Against its own pre-trained version the cuts were 7.7% and 17.4%.
On uncalibrated smartphone inertial data, the authors say the method limits drift and keeps the path aligned despite noisy sensors. The code is planned for release.
Background
DFIO
Inertial odometry method
A learning method that builds a Kalman filter update into training to reduce drift in motion tracking from inertial sensors.
