Rigid-body trajectory datasets

Content overview

This repository consists of two rigid-body trajectory datasets DLA and SYN.DLA is an existing datasets and SYN is a newly developed dataset.  Data-handing efforts have been made to resample the rigid-body trajectories towards a fixed sample frequency of 50Hz, and to format all the data to consistent .csv format.

For each dataset, every trial_x.csv file is a Comma-Separated Values (CSV) file. The trailing number x refers to the order in which the trials were performed. The file trial_x.csv has the following columns:

The first column represents the time axis, consisting of the time stamps at a sampling frequency of 50Hz.

The second to fourth columns contain the xyz-position coordinates of the origin of the body's reference frame. 

The fifth to eighth columns contain the quaternion coordinates of the orientation of the body's reference frame. The quaternion coordinates adhere to the scalar first convention.

Description of the Daily Life Activities (DLA) dataset

The first dataset is the Daily Life Activities (DLA) dataset retrieved from link. This dataset consists of trials of daily-life object-manipulation tasks performed by a human. The dataset consists of ten motion classes: cutting, painting, pouring with cup, putting cup away, quarter turn, scooping and pouring, scooping food, shaking, sinusoidal motion, and table wiping.

In this dataset, a high variation in the context was purposefully introduced. That is, the motions were performed with respect to three different viewpoints (V1, V2, V3) and with four different execution styles (normal, with larger spatial scale, with different velocity profile, and with longer time duration). This resulted in a total of (3x4=12) twelve different contexts in which the motions were performed. Each motion class was performed ten times in every context, resulting in a total of (10x3x4x10=1200) measured trials.

As explained in link, the trials were recorded using a Krypton K600 camera from NIKON Metrology by tracking up to twelve LED markers attached to the manipulated object. The 3D position of each LED marker was recorded with a sampling rate of 50Hz and an expected accuracy of 0.4mm. A selection of the data was made by only considering the trials that have at least three visible LED markers during the entire motion, since, in that case, the rigid-body motion can be uniquely determined. After this process, 1079 out of the 1200 measured trials remained.

For the remaining trials, the rigid-body trajectory (position + quaternion coordinates) of the manipulated object was extracted from the measured LED marker trajectories. The average position of the LED markers served as the object reference frame origin. The object reference frame orientation was extracted from the positions of three LED markers using a Gram-Schmidt orthonormalisation algorithm. 

Citing

If you use this DLA dataset, please cite it as follows:

@INPROCEEDINGS{... ,
author={Vochten, Maxim and De Laet, Tinne and De Schutter, Joris},
booktitle={2015 IEEE International Conference on Robotics and Automation (ICRA)},
title={Comparison of rigid body motion trajectory descriptors for motion representation and recognition},
year={2015},
volume={},
number={},
pages={3010-3017},
doi={10.1109/ICRA.2015.7139612}}

Description of the Synthetic rigid-body motion (SYN) dataset

The second dataset is the newly developed Synthetic rigid-body motion (SYN) dataset. The SYN dataset was purposefully designed to test the performance of different invariant trajectory descriptors when applied to elementary rigid-body motions. This dataset consists of seven rigid-body motion classes:

Axis: rotation about a stationary axis

Negative_screw: screw motion with a negative pitch

Positive_screw: screw motion with a positive pitch

precession: precession motion

Trans: translation along a straight line

Trans_circle: circular translation

Trans_helix: helical translation

The trajectories of each motion class were simulated in three different contexts:

Original: The original position and orientation of both the world and body reference frames were maintained.

Change_in_references_1 and Change_in_references_2: The original position and orientation of both the world and body reference frames were displaced.

For every context, multiple trials were synthetically generated (four for the original context and 99 for the other contexts). Brown noise was added to the trajectory data, simulating lower frequency disturbances as if the motions were performed by humans and measurement noise was present. In total, this resulted in 7 x ( 4 + 99 + 99) = 1414 trials.

Citing

If you use this SYN dataset, please cite it as follows:

@misc{...,     title={BILTS: A Bi-Invariant Similarity Measure for Robust Object Trajectory Recognition under Reference Frame Variations},       author={Arno Verduyn and Erwin Aertbeliën and Glenn Maes and Joris De Schutter and Maxim Vochten},     year={2025},      eprint={2405.04392},      archivePrefix={arXiv},      primaryClass={cs.RO},      url={https://arxiv.org/abs/2405.04392}, }