RoNIN-based PDR for SensorLogger

This fork modifies RoNIN to be used with data (acc, gyro, rotvec) collected with SensorLogger data collection app (Android and iOS). The fork's main purpose is to offer an easy RoNIN-based PDR algorithm for researchers and developers, as I found it to be surprisingly hard to find robust and working PDR options freely available. Or better yet, usable out of the box.

The compatibility is achieved by hacking RoNIN (ridi execution path) internals step by step, removing blocking dependencies on ground truth and calibration data.

Setup

1. Clone this fork
2. Download a pre-trained RoNIN model you want to use (from here). The fork is tested with ronin_resnet/checkpoint_gsn_latest.pt.
3. Create virtual environment, activate, and install dependencies (or run globally)

python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt

Usage

Example notebook at sensorlogger_to_pdr/example.ipynb demonstrates basic usage.

Command line:

python ./sensorlogger/sensorlogger_to_pdr.py --path <sensor_logger_data> --ronin_path <ronin_path> --model_path <ronin_model_path>

Example:

python ./sensorlogger/sensorlogger_to_pdr.py --path /ronin_for_sensorlogger/sensorlogger/example_data/shopping_mall_eight --ronin_path "/ronin_for_sensorlogger/" --model_path /ronin_for_sensorlogger/models/ronin_resnet/checkpoint_gsn_latest.pt

Sensor and data requirements

The modified RoNIN algorithm requires the following sensor data files

• Gyroscope.csv
• Orientation.csv
• TotalAcceleration.csv

Using other data collection applications should be straightforward by just converting to SensorLogger format or converting directly to the pickle RoNIN requires.

Gyroscope.csv format:

time,seconds_elapsed,z,y,x
1718030736872695300,0.0706953125,-0.32157647609710693,0.15512901544570923,0.6729803085327148

Orientation.csv format:

time,seconds_elapsed,qz,qy,qx,qw,roll,pitch,yaw
1718030736912338000,0.110337890625,-0.49317896366119385,-0.10824661701917648,0.21374563872814178,0.8362834453582764,0.033629775047302246,-0.4828145205974579,1.0739439725875854

NOTE: roll, pitch, and yaw are not required by RoNIN

TotalAcceleration.csv format:

time,seconds_elapsed,z,y,x
1716979893874983200,0.047983154296875,9.851848602294922,1.407876968383789,-0.008374879136681557

Notes:

The fork was originally hacked under a timeline for research purposes. The functionality has not been fully tested, but the fork seems to work over different phone models and manufacturers (OnePlus, iPhone). Fixes and suggestions are welcome.

Original README:

RoNIN: Robust Neural Inertial Navigation in the Wild

Paper: ICRA 2020, arXiv
Website: http://ronin.cs.sfu.ca/
Demo: https://youtu.be/JkL3O9jFYrE

---

Requirements

python3, numpy, scipy, pandas, h5py, numpy-quaternion, matplotlib, torch, torchvision, tensorboardX, numba, plyfile, tqdm, scikit-learn

Data

The dataset used by this project is collected using an App for Google Tango Device and an App for any Android Device, and pre_processed to the data format specified here Please refer to our paper for more details on data collection.

You can download the RoNIN dataset from our project website or HERE. Unfortunately, due to security concerns we were unable to publish 50% of our dataset.

Optionally, you can write a custom dataloader (E.g: soure/data_ridi.py) to load a different dataset.

Usage:

1. Clone the repository.
2. (Optional) Download the dataset and the pre-trained models¹ from HERE.
3. Position Networks
   1. To train/test RoNIN ResNet model:
      • run source/ronin_resnet.py with mode argument. Please refer to the source code for the full list of command line arguments.
      • Example training command: python ronin_resnet.py --mode train --train_list <path-to-train-list> --root_dir <path-to-dataset-folder> --out_dir <path-to-output-folder>.
      • Example testing command: python ronin_resnet.py --mode test --test_list <path-to-train-list> --root_dir <path-to-dataset-folder> --out_dir <path-to-output-folder> --model_path <path-to-model-checkpoint>.
   2. To train/test RoNIN LSTM or RoNIN TCN model:
      • run source/ronin_lstm_tcn.py with mode (train/test) and model type. Please refer to the source code for the full list of command line arguments. Optionally you can specify a configuration file such as config/temporal_model_defaults.json with the data paths.
      • Example training command: python ronin_lstm_tcn.py train --type tcn --config <path-to-your-config-file> --out_dir <path-to-output-folder> --use_scheduler.
      • Example testing command: python ronin_lstm_tcn.py test --type tcn --test_list <path-to-test-list> --data_dir <path-to-dataset-folder> --out_dir <path-to-output-folder> --model_path <path-to-model-checkpoint>.
4. Heading Network
   • run source/ronin_body_heading.py with mode (train/test). Please refer to the source code for the full list of command line arguments. Optionally you can specify a configuration file such as config/heading_model_defaults.json with the data paths.
   • Example training command: python ronin_body_heading.py train --config <path-to-your-config-file> --out_dir <path-to-output-folder> --weights 1.0,0.2.
   • Example testing command: python ronin_body_heading.py test --config <path-to-your-config-file> --test_list <path-to-test-list> --out_dir <path-to-output-folder> --model_path <path-to-model-checkpoint>.

¹ The models are trained on the entire dataset

Citation

Please cite the following paper is you use the code, paper or data:
Herath, S., Yan, H. and Furukawa, Y., 2020, May. RoNIN: Robust Neural Inertial Navigation in the Wild: Benchmark, Evaluations, & New Methods. In 2020 IEEE International Conference on Robotics and Automation (ICRA) (pp. 3146-3152). IEEE.