Datasets

[Mayitzin]

A great number of datasets can be found online, and I'm permanently looking for good ones. I'd recommend to use the ones including:

• Gyroscope data.
• Accelerometer data.
• Magnetometer data.
• Ground truth of orientations.

Some may lack the data of one of the sensors. Some others might have even more data than you expect (like videos or images for visual dometry.)

Check if you can use them to test your applications. Here I will post some of the ones I've found useful, their pros and cons. Maybe you can get some nice results with them.

[Mayitzin]

RepoIMU

Location: https://github.com/agnieszkaszczesna/RepoIMU
Author: Agnieszka Szczesna, Przemysław Skurowski, Przemysław Pruszowski, Damian Pęszor, Marcin Paszkuta, and Konrad Wojciechowski.
Documentation: https://www.researchgate.net/publication/307076277

Is the first dataset that comes to mind. It has full 9-dof recordings at 100 Hz, synchronized to their ground truths as quaternions. One of the files of this dataset was actually used for initial tests of AHRS, and you might find it in the tests folder.

Pros:

• Full data of the 3 types of main sensors.
• Ground truth (as quaternions).
• Sensor data and ground truth are already synchronized at 100 Hz.
• Includes static recordings to analyze stabilty of sensor signals.
• Single CSV File per recording with full data.

Cons:

• Recordings do not reflect real user applications, only controlled angular motions.
• Does not include translation (position of sensor frame w.r.t. the global frame.)

I personally would use this dataset to understand the angular motion, and test drift after angular integration.

[Mayitzin]

I moved this topic to issue #31 because I think it is more convenient to discuss it there, as this discussion was created precisely to be about the RepoIMU dataset.

For further discussion about the transformation of quaternion to roll-pitch-yaew angles, please go there.

[PeterBorisenko]

My personal observation is that this dataset seems to provide deviated magnetometer data.
The best results I could get was with simple gyroscope integration. Any use of magnetometer data makes a mess.

[PeterBorisenko]

I would like to know as well which variance for each sensor and which magnetic reference do you use with this data set in case of EKF?

[Mayitzin]

Attitude Estimation with Smartphones

• Location: https://github.com/tyrex-team/benchmarks-attitude-smartphones
• Author: Michel Thibaud, Hassen Fourati, Pierre Genevès, Nabil Layaïda
• Documentation: http://tyrex.inria.fr/mobile/benchmarks-attitude/
• License: CeCILL

This dataset is focused on smartphones, and their recordings are obtained from their embedded IMUs. Three smartphones with different sensor models are used:

Model	Accelerometer	Gyroscope	Magnetometer
iPhone 4S	STMicro STM33DH (100 Hz)	STMicro AGDI (100 Hz)	AKM 8975 (40 Hz)
iPhone 5	STMicro LIS331DLH (100 Hz)	STMicro L3G4200D (100 Hz)	AKM 8963 (25 Hz)
Nexus 5	InvenSense MPU6515 (200 Hz)	InvenSense MPU6515 (200 Hz)	AKM 8963 (60 Hz)

Eight typical motions are recorded:

1. Augmented Reality.
2. Texting while walking.
3. Phoning while walking.
4. Swinging hand while walking.
5. In the front pocket while walking.
6. In the back pocket while walking.
7. In the front pocket while running.
8. In the hand while running.

Pros:

• Very practical use cases.
• Includes recordings with and without calibration of gyroscopes and magnetometers.
• Includes positions of sensor frame w.r.t. global frame.
• Ground truths are provided as rotation matrices and as roll-pitch-yaw angles.

Cons:

• Data is not synchronized.
• Some ground truth files have "jumps" with no data, due to the markers being covered during the recording.
• Files are in MAT format, which makes it difficult to read and organize.

[anisghaoui]

OxIOD: The Dataset for Deep Inertial Odometry

---

Location: http://deepio.cs.ox.ac.uk/
Authors: Changhao Chen, Peijun Zhao, Chris Xiaoxuan Lu, Wei Wang, Andrew Markham, Niki Trigoni
Date: 2018
Size: 1 GB
Duration: 14.72 hours total
Documentation: not found

This dataset is delivered with smartphone's IMU, Vicon and Google Tango phone (Visual-inertial SLAM). It was created to train their neural network.

Various phone positions and usages:

1. handbag
2. handheld
3. pocket
4. trolley
5. running
6. slow walking
7. large scale
8. multi devices
9. multi users

Pros:

• Raw and synchronised data.
• All in .csv
• Both ground truths for small experiments and large scale
• frequency = 100 Hz

Cons:

• sync data have their timestamps overwritten as 1.5E11. You can still recover the correct timestamping.

[Mayitzin]

Berlin Robust Orientation Estimation Assessment Dataset (BROAD)

• Location: https://github.com/dlaidig/broad
• Authors: D. Laidig, M. Caruso, A. Cereatti, T. Seel.
• Documentation: https://www.mdpi.com/2306-5729/6/7/72/pdf
• License: Creative Commons BY 4.0

The dataset includes 3D inertial and magnetic data with synchronized optical marker-based ground truth measurement. It consists of 39 trials at different speeds and various types of movement. Thereof, 23 trials are performed in an undisturbed indoor environment, and 16 trials are recorded with deliberate magnetometer and accelerometer disturbances.

• IMU data recorded at 286 Hz using a nine-axis inertial sensor myon aktos-t.
• Ground truth data at 120 Hz was obtained via an Optitrack OMC system consisting of eight Flex13 cameras.

Data in each file is:

• imu_gyr IMU gyroscope measurements in rad/s.
• imu_acc IMU accelerometer measurements in m/s2.
• imu_mag IMU magnetometer measurements in μT.
• opt_quat OMC ground truth orientation as a unit quaternion (w component first, ENU reference frame).
• opt_pos OMC ground truth position in m.
• movement Boolean array (0/1) indicating movement phases.
• sampling_rate sampling rate of the measurements in Hz (2000/7 Hz ≈ 286 Hz, HDF5: stored as an attribute).
• info short description of the file contents (HDF5: stored as an attribute).

Pros:

• Different environments and cases.
• Ground truths are provided as quaternions avoiding any gimbal lock problems and reducing the processed data.

Cons:

• Files are in either HD5 or MAT format, which makes it difficult to read and organize. Extra tools are required to read the dataset (provided in its repository)