Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AddBiomechanics Dataset: Capturing the Physics of Human Motion at Scale

Keenon Werling, Janelle Kaneda, Alan Tan, Rishi Agarwal, Six Skov, Tom Van Wouwe, Scott Uhlrich, Nicholas Bianco, Carmichael Ong, Antoine Falisse, Shardul Sapkota, Aidan Chandra, Joshua Carter, Ezio Preatoni, Benjamin Fregly, Jennifer Hicks, Scott Delp, C. Karen Liu

TL;DR

This work presents the AddBiomechanics Dataset 1.0, which includes physically accurate human dynamics of 273 human subjects, over 70 hours of motion and force plate data, totaling more than 24 million frames, and proposes a benchmark for estimating human dynamics from motion using this dataset.

Abstract

While reconstructing human poses in 3D from inexpensive sensors has advanced significantly in recent years, quantifying the dynamics of human motion, including the muscle-generated joint torques and external forces, remains a challenge. Prior attempts to estimate physics from reconstructed human poses have been hampered by a lack of datasets with high-quality pose and force data for a variety of movements. We present the AddBiomechanics Dataset 1.0, which includes physically accurate human dynamics of 273 human subjects, over 70 hours of motion and force plate data, totaling more than 24 million frames. To construct this dataset, novel analytical methods were required, which are also reported here. We propose a benchmark for estimating human dynamics from motion using this dataset, and present several baseline results. The AddBiomechanics Dataset is publicly available at https://addbiomechanics.org/download_data.html.

AddBiomechanics Dataset: Capturing the Physics of Human Motion at Scale

TL;DR

This work presents the AddBiomechanics Dataset 1.0, which includes physically accurate human dynamics of 273 human subjects, over 70 hours of motion and force plate data, totaling more than 24 million frames, and proposes a benchmark for estimating human dynamics from motion using this dataset.

Abstract

While reconstructing human poses in 3D from inexpensive sensors has advanced significantly in recent years, quantifying the dynamics of human motion, including the muscle-generated joint torques and external forces, remains a challenge. Prior attempts to estimate physics from reconstructed human poses have been hampered by a lack of datasets with high-quality pose and force data for a variety of movements. We present the AddBiomechanics Dataset 1.0, which includes physically accurate human dynamics of 273 human subjects, over 70 hours of motion and force plate data, totaling more than 24 million frames. To construct this dataset, novel analytical methods were required, which are also reported here. We propose a benchmark for estimating human dynamics from motion using this dataset, and present several baseline results. The AddBiomechanics Dataset is publicly available at https://addbiomechanics.org/download_data.html.

Paper Structure

This paper contains 16 sections, 10 equations, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Related Datasets
  4. Estimating Physical Forces from Motion
  5. AddBiomechanics Dataset
  6. Personalized Subject Models
  7. Evaluating Dataset Quality
  8. Raw Data Processing and Aggregation
  9. Processing Raw Sensor Data
  10. Fitting Continuous Dynamics While Ignoring Frames with Missing Force Data
  11. Ablation Study
  12. Benchmark for Estimating Dynamics from Motion
  13. Task Description
  14. Proposed Metrics
  15. Preliminary Baseline Experiments
  16. ...and 1 more sections

Figures (6)

  • Figure 1: Frames from recordings of a drop jump, tai chi, running, ballet, and squatting (left to right) in the AddBiomechanics Dataset. The ground reaction forces are shown with arrows on the right (orange) and left (blue) feet for representative time steps around the shown frame. The model includes 22 body segments, 37 degrees of freedom, and specialized joints for the knees, shoulders, and spine, which match the subject's mass, inertial properties, and motion capture recordings while obeying $F=ma$ with respect to the force plate recordings. Discrepancies between the reconstruction and the original sensor data are within clinically acceptable tolerances hicks2015my.
  • Figure 2: Activity classification. The duration of captures in each activity class is shown on a log scale.
  • Figure 3: Distributions of speed and capture length in the AddBiomechanics Dataset. The magnitude of speed averaged per trial (left). Trial lengths as number of frames (right). Original trials longer than 2000 frames were split up into multiple trials segments.
  • Figure 4: Evaluating dataset quality. See Table \ref{['tab:hicks-threshold']} for descriptions of quantities.
  • Figure 5: Adapted with permission from Werling et. al. werling2023addbiomechanics, this figure shows the processing pipeline we used to turn raw sensor data into accurate human dynamics, and shows where the method described below swaps into the original pipeline from werling2023addbiomechanics
  • ...and 1 more figures