PyVHR
Python framework for Virtual Heart Rate
Install / Use
/learn @phuselab/PyVHRREADME
Package pyVHR (short for Python framework for Virtual Heart Rate) is a comprehensive framework for studying methods of pulse rate estimation relying on video, also known as remote photoplethysmography (rPPG).
pyVHR: a Python framework for remote photoplethysmography
Description
The methodological rationale behind the framework is that in order to study, develop and compare new rPPG methods in a principled and reproducible way, the following conditions should be met: i) a structured pipeline to monitor rPPG algorithms' input, output, and main control parameters; ii) the availability and the use of multiple datasets; iii) a sound statistical assessment of methods' performance. pyVHR allows to easily handle rPPGmethods and data, while simplifying the statistical assessment. Its main features lie in the following:
- Analysis-oriented. It constitutes a platform for experiment design, involving an arbitrary number of methods applied to multiple video datasets. It provides a systemic end-to-end pipeline, allowing to assess different rPPG algorithms, by easily setting parameters and meta-parameters.
- Openness. It comprises both method and dataset factory, so to easily extend the pool of elements to be evaluatedwith newly developed rPPG methods and any kind of videodatasets.
- Robust assessment. The outcomes are arranged intostructured data ready for in-depth analyses. Performance comparison is carried out based on robust nonparametric statistical tests.
Nine classical rPPG methods, namely ICA, PCA, GREEN, CHROM, POS, SSR, LGI, PBV, OMIT, as well as the recent Deep Learning-based model MTTS-CAN are implemented. Moreover, pyVHR provides APIs for handling 11 publicly available video datasets, i.e. PURE, LGI-PPGI-DB, UBFC-1, , UBFC-2, UBFC-Phys, ECG-Fitness, MAHNOB Vicar-PPG-2, V4V , VIPL-HR and COHFACE, usually adopted to benchmark rPPG methods. Eventually, extensive rigorous statistical analyses can be effortlessly performed via the pyVHR stats APIs.
Getting started
Dependencies
The quickest way to get started is to install the miniconda distribution, a lightweight minimal installation of Anaconda Python.
Once installed, create a new conda environment and automatically fetch all the dependencies based on your architecture (with or without GPU), using one of the following commands:
CPU-only version (v. 1.2 - previous version)
conda env create --file https://raw.githubusercontent.com/phuselab/pyVHR/pyVHR_CPU/pyVHR_CPU_env.yml
CPU+GPU version (v. 2.0 - current version)
This yml environment is for cudatoolkit=11.3 and python=3.9.
conda env create --file https://raw.githubusercontent.com/phuselab/pyVHR/master/pyVHR_env.yml
Installation
Enter the newly created conda environment and install the latest stable release build of pyVHR with:
CPU-only version (v. 1.2 - previous version)
conda activate pyvhr
(pyvhr) pip install pyvhr-cpu
CPU+GPU version (v. 2.0 - current version)
conda activate pyvhr
(pyvhr) pip install pyvhr
Basic usage
Run the following code to obtain BPM estimates over time for a single video:
from pyVHR.analysis.pipeline import Pipeline
from pyVHR.plot.visualize import *
from pyVHR.utils.errors import getErrors, printErrors, displayErrors
# params
wsize = 6 # window size in seconds
roi_approach = 'patches' # 'holistic' or 'patches'
bpm_est = 'clustering' # BPM final estimate, if patches choose 'medians' or 'clustering'
method = 'cpu_CHROM' # one of the methods implemented in pyVHR
# run
pipe = Pipeline() # object to execute the pipeline
bvps, timesES, bpmES = pipe.run_on_video(videoFileName,
winsize=wsize,
roi_method='convexhull',
roi_approach=roi_approach,
method=method,
estimate=bpm_est,
patch_size=0,
RGB_LOW_HIGH_TH=(5,230),
Skin_LOW_HIGH_TH=(5,230),
pre_filt=True,
post_filt=True,
cuda=True,
verb=True)
# ERRORS
RMSE, MAE, MAX, PCC, CCC, SNR = getErrors(bvps, fps, bpmES, bpmGT, timesES, timesGT)
printErrors(RMSE, MAE, MAX, PCC, CCC, SNR)
displayErrors(bpmES, bpmGT, timesES, timesGT)
The full documentation of run_on_video method, with all the possible parameters, can be found here: https://phuselab.github.io/pyVHR/
Notebooks
Some demonstration jupyter notebooks that help to better understand the many features of the framework are contained in the notebooks folder.
pyVHR_demo.ipynb: Basic demo with individual steps explained in detail.pyVHR_run_on_video.ipynb: Show execution on a single video by deriving HRVs and error values from the reference signal.pyVHR_run_on_dataset.ipynb: Show execution on a single dataset by deriving HRVs and error values from the reference signals. It is also possible to make some basic statistics, boxplots and ranking tests for comparative purposes.pyVHR_demo_deep.ipynb: Show execution of deep methods on a single dataset by deriving HRV and error values from reference signals.
Documentation
The full documentation of the pyVHR framework is available at https://phuselab.github.io/pyVHR/.
Developing
The latest unstable development build of pyVHR is available on GitHub, and can be obtained downloading from source and installing via:
git clone git@github.com:phuselab/pyVHR.git
cd pyVHR/
python setup.py install
The main branch refers to the full pyVHR framework (requires GPU), while the pyVHR_CPU branch is dedicated to the CPU-only architectures.
Custom installation
If you want to create your environment from scratch you should follow these steps:
- Install PyTorch (here)
- Install Numba (here)
- Install Cupy (for GPU only) with the correct CUDA version (here)
- Install CuSignal (for GPU only) using conda and remove from the command 'cudatoolkit=x.y' (here)
- Install Kaleido (here)
- Install PyTables (here)
- Install pyVHR as shown above.
Methods
The framework contains the implementation of many common methods for remote-PPG measurement. Currently implemented methods with reference publications are:
| Method name | Reference paper | | ------------ | ---------------------------------------------------------------------- | |Green | Verkruysse, W., Svaasand, L. O., & Nelson, J. S. (2008). Remote plethysmographic imaging using ambient light. Optics express, 16(26), 21434-21445.| |CHROM | De Haan, G., & Jeanne, V. (2013). Robust pulse rate from chrominance-based rPPG. IEEE Transactions on Biomedical Engineering, 60(10), 2878-2886.| |ICA | Poh, M. Z., McDuff, D. J., & Picard, R. W. (2010). Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Optics express, 18(10), 10762-10774.| |LGI | Pilz, C. S., Zaunseder, S., Krajewski, J., & Blazek, V. (2018). Local group invariance for heart rate estimation from face videos in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (pp. 1254-1262).| |PBV | De Haan, G., & Van Leest, A. (2014). Improved motion robustness of remote-PPG by using the blood volume pulse signature. Physiological measurement, 35(9), 1913.| |PCA | Lewandowska, M., Rumiński, J., Kocejko, T., & Nowak, J. (2011, September). Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In 2011 federated conference on computer science and information systems (FedCSIS) (pp. 405-410). IEEE.| |POS | Wang, W., den Brinker, A. C., Stuijk, S., & de Haan, G. (2016). Algorithmic principles of remote PPG. IEEE Transactions on Biomedical Engineering, 64(7), 1479-1491.| |SSR | Wang, W., Stuijk, S., & De Haan, G. (2015). A novel algorithm for remote photoplethysmography: Spatial subspace rotation. IEEE transactions on biomedical engineering, 63(9), 1974-1984.| |OMIT | Álvarez Casado, C., Bordallo López, M. (2022). Face2PPG: An unsupervised pipeline for blood volume pulse extraction from faces. arXiv (eprint 2202.04101).| |MTTS-CAN | Liu, X., Fromm, J.,
