Deep IMU Processing

Deep Learning for Inertial Positioning: A Survey 2023 forces on 2018-2022 updates.

• Pedestrian : incorporating domainspecific knowledge or other sensor. (1) pedestrian dead reckoning (PDR): detecting steps, estimating step length, heading. (2) zero-velocity update (ZUPT).
• Inertial sensor calibration. (typically dependent on the specific sensor or platform used)
  1. Supervised by more accurate IMU data (expensive).
  2. Supervised by integrated orientation. OriNet 2019. Calib-Net 2022.
  3. Learn calibration parameters. Learning to calibrate 2019 uses RL.
• Imu integration correction.
  1. Learn location displacement - the average velocity multiplied by a fixed period of time.
     • IONet 2018 the frequency of platform vibrations - absolute moving speed.
     • MotionTransformer 2019 extend adaptability using GAN.
     • TLIO 2020 extend to 3d pedestrian, use EKF.
  2. Learn velocity - to correct accelerations. mostly in pedestrian.
  3. Learn velocity - to correct KF state.
     • AI-IMU 2020 estimate velocity covariance.
     • RNN 2021 Alibaba, model motion with imu only.
     • DeepVIP 2022 (indoor) velocity & heading.
     • OdoNet 2022 speed learning and ZUPT.
• Correct pedestrian position, learn PDR and learn ZUPT.
• Deep sensor fusion, mostly VIO.
  • Supervised end-to-end.
  • Unsupervised using view synthesis.

OdoNet: Untethered Speed Aiding for Vehicle Navigation Without Hardware Wheeled Odometer 2021, CNN sequence IMU to predict front velocity.

• make a full system to combine CNN with filter.
  • CNN module + INS module + Filter Fusion module.
  • consider also the car-imu extrinsics, and the IMU noise model (fixed by Filter).
• use CNN instead of RNN, for fast convergence and better accuracy.
• We cannot tell the difference between static car and constant velocity motion, so it should be better to estimate delta velocity.

Inertial sensing meets machine learning: Opportunity or challenge? 2020 similar structure as the more latest paper, but includes more old researches.

AI-IMU Dead-Reckoning 2020, github. predict the front velocity covariance, then use the constraint for state estimation. I have been testing IMU-positioning based on this model.