Awesome
RIDI: Robust IMU Double Integration
Introducing RoNIN: Robust Inertial Neural Navigation
We have developed a far superior algorithm based on deep learning, which is simpler, more accurate and more generalizable:
RoNIN: Robust Nueral Inertial Navigation http://ronin.cs.sfu.ca
We highly recommend you try the new algorithm instead of RIDI. It is still beneficial to read the RIDI paper for our initial idea of data-driven inertial navigation, but I would NOT suggest you dive too deep into the code.
Prerequisite
- Python: numpy, scipy, opencv-python (>3.0), numpy-quaternion, plyfile
- C++: Glog, Gflags, OpenCV (>3.0), Eigen (>3.3), Ceres-Solver, OpenMesh
Data format
The dataset used by this project is collected by a specific Tango App: https://github.com/higerra/TangoIMURecorder Please read the README.md file of this repository for detailed file format. If you do not have a Tango phone, you can use the alternative app: https://github.com/higerra/AndroidIMURecorder Note that without a Tango phone, the ground truth trajectory will not be available.
Usage:
- Clone the repository.
- (Optional) Download the dataset from HERE and the pre-trained model from HERE. Note that the pre-trained model is trained from a small group of people. For the best result please train your own model.
- For newly captured dataset, run
python/gen_dataset.py(with Tango phone) orpython/gen_dataset_nopose.py(without Tango phone) to preprocess the dataset. Please refer to the source code for command line arguments. - To train a model, run
python/regression_cascade.py. Please refer to the source code for command line arguments. One possible call is: *python/regression_cascade.py --list <path-to-dataset-list> --model_output <path-to-output-folder>. - Compile the C++ source code:
*
cd <project-root>*mkdir cpp/build & cd cpp/build*cmake ..*make -j4 - After the compilation finishes, run the following executable to perform RIDI:
*
cd <project-root>/cpp/build/imu_localization*./IMULocalization_cli <path-to-dataset> --model_path <path-to-model> - The result will be written to
<path-to-dataset>/result_full
Citation
Please cite the following paper is you use the code: