Annonce en lien avec l’Action/le Réseau : aucun
Laboratoire/Entreprise : ImViA – Univ. Bourgogne
Durée : 36 months
Contact : yannick.benezeth@u-bourgogne.fr
Date limite de publication : 2019-05-15
Contexte :
Video analysis can be used to recover subtle changes in the human body associated with health parameters. Non-contact video-based physiological measurement is a very active research domain with great applications. Recently, several algorithms have been proposed to enable recovering vital signs from only a webcam video with limited success under challenging conditions (NIR monitoring, motion, etc.). The technics are usually based on blind source separation paradigms (e.g., constrained ICA [2]) or on simple color space transforms based on the light-skin interaction models (e.g., CHROM [3], POS, etc.). In this work, we propose to investigate supervised learning technics for video-based measurement of physiological parameters, including deep recurrent networks with an end-to-end strategy.
Sujet :
In computer vision, end-to-end deep neural models have out-performed traditional multi-stage methods that require hand-crafted feature manipulation. An end-to-end learning framework for recovering physiological signals would be highly desirable. Currently, the technics are based on blind source separation paradigms, or color space transforms based on the light-skin interaction model. However, these methods still struggle to face specific challenges, especially motion disturbances. Recently, a CNN network has been proposed to tackle this problem [1]. Even if this work has proved the feasibility of the approach, many works remain for practical applications. We plan to primarily investigate the inclusion of motion disturbances in the network and apply dedicated strategies to allow long term memories management in the recurrent network.
We have been working on remote photoplethysmography for 4 years now and have contributed to the two existing strategies [4, 5, 6, 7] (blind source separation or color space transforms). It now seems very important to explore this new avenue of research in the field. Moreover, over the years, we have collected a large amount of data that can be used for the learning of the models.
Supervision team:
Johel Miteran (miteranj@u-bourgogne.fr)
Yannick Benezeth (yannick.benezeth@u-bourgogne.fr) – http://sites.google.com/view/ybenezeth
Application
Send an email (before May 15th, 2019) to J. Miteran and Y. Benezeth with:
– A motivation letter
– Official transcripts (Master and Bachelor)
– CV
– An example of research work (paper, Master thesis, etc.)
– Recommendations or reference persons to contact
References
[1] W. Chen and D. McDuff, “DeepPhys: Video-Based Physiological Measurement Using Convolutional Attention Networks,” in ECCV, 2018.
[2] R. Macwan, Y. Benezeth, A. Mansouri. Heart rate estimation using remote photoplethysmography with multi-objective optimization. Biomedical Signal Processing and Control, Elsevier, 2019
[3] G. De Haan and V. Jeanne, “Robust pulse rate from chrominance-based rPPG,” IEEE Trans. Biomed. Eng., vol. 60, no. 10, pp. 2878–2886, 2013.
[4] R. Macwan, Y. Benezeth, A. Mansouri. Heart rate estimation using remote photoplethysmography with multi-objective optimization. Biomedical Signal Processing and Control, Elsevier, 2019
[5] S. Bobbia, D. Luguern, Y. Benezeth, K. Nakamura, R. Gomez and J. Dubois, “Real-Time Temporal Superpixels for Unsupervised Remote Photoplethysmography”, The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018.
[6] R. Macwan, S. Bobbia, Y. Benezeth, J. Dubois, A. Mansouri, “Periodic Variance Maximization using Generalized Eigenvalue Decomposition applied to Remote Photoplethysmography estimation”, The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018.
[7] S. Bobbia, R. Macwan, Y. Benezeth, A. Mansouri, J. Dubois, Unsupervised skin tissue segmentation for remote photoplethysmography, Pattern Recognition Letters, Elsevier, 2017.
Profil du candidat :
Required background:
– Computer Vision or Signal Processing
– Experiences in using recent supervised learning frameworks (Keras, TensorFlow, …) would be appreciated.
Formation et compétences requises :
Master degree in Computer Vision, Signal Processing or Machine learning
Adresse d’emploi :
Dijon France – Univ. Bourgogne