Robust Real-Time Pulse Rate Estimation From FacialVideo Using Sparse Spectral Peak Tracking
Aditya Gaonkar P1 Bhuthesh R2 Dipanjan Gope2 PrasantaKumar Ghosh1
1Department of Electrical EngineeringIndian Institute of Science
2Department of Electrical Communication EngineeringIndian Institute of Science
SPCOM, 2016
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 1 / 25
A challenge!
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 2 / 25
Outline
1 Motivation
2 Data collection
3 Proposed Algorithm
4 Experiments and Results
5 Conclusions
6 Future work
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 3 / 25
Motivation
Outline
1 Motivation
2 Data collection
3 Proposed Algorithm
4 Experiments and Results
5 Conclusions
6 Future work
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 4 / 25
Motivation
Motivation for this study
(a) A person shouting at his smartphone (b) A typical example of telemedicine
Situations where remotely monitoring pulse rate can be beneficial
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 5 / 25
Motivation
Summary of this work
Prior methods
Prior methods for computing pulse rate from facial video include ICAbased techniques, Machine Learning based techniques etc.
In all these works, the pulse rate is computed independently in eachanalysis window.
The performance on our dataset obtained by using the state of theart methods was 14.62 beats per minute (BPM) in root mean squareerror (RMSE).
Our contribution
This work explores benefits of constraining the change of the PRvalue between successive analysis windows.
We have obtained an improvement of 6.71 BPM in RMSE on ourdataset over the state of the art methods.
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 6 / 25
Data collection
Outline
1 Motivation
2 Data collection
3 Proposed Algorithm
4 Experiments and Results
5 Conclusions
6 Future work
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 7 / 25
Data collection
Recording setup
The resolution of the videos were 1280x720p.
The Samsung phone’s videos were in mp4 format and iPhone’s videoswere in avi format.
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 8 / 25
Data collection
The FaViP dataset
This dataset is available for download in our lab webpage 1.S U B J E C T 1 S U B J E C T 2 S U B J E C T 3 S U B J E C T 4 S U B J E C T 5
S U B J E C T 6 S U B J E C T 7 S U B J E C T 8 S U B J E C T 9 S U B J E C T 10
S U B J E C T 11 S U B J E C T 12 S U B J E C T 13 S U B J E C T 14 S U B J E C T 15
A good amount of variability across the subjects is illustrated.
1at http://spire.ee.iisc.ernet.in/spire/database.phpAditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 9 / 25
Proposed Algorithm
Outline
1 Motivation
2 Data collection
3 Proposed Algorithm
4 Experiments and Results
5 Conclusions
6 Future work
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 10 / 25
Proposed Algorithm
Motivation
In all prior works, decision is taken independently in each analysiswindow.
Pulse rate is a slowly varying quantity and for 2 analysis windowsalmost overlapping each other, pulse rate wouldnt have changed by ahuge margin.
The above fact has been exploited in the proposed algorithm.
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 11 / 25
Proposed Algorithm
State of the art method
The above block diagram illustrates the method proposed inthe state of the art method a
aMing-Zher Poh, Daniel J McDuff, and Rosalind W Picard, “Non-contact,automated cardiac pulse measurements using video imaging and blind sourceseparation”, Optics express, vol. 18, no. 10, pp. 1076210774, 2010.
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 12 / 25
Proposed Algorithm
Illustrating the state of the art method
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 13 / 25
Proposed Algorithm
Illustrating the state of the art method
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 13 / 25
Proposed Algorithm
Illustrating the state of the art method
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 13 / 25
Proposed Algorithm
Illustrating the state of the art method
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 13 / 25
Proposed Algorithm
Additional step #1: Spectral windowing of the ICA traces
We propose spectral windowing of the ICA traces in eachsegment using various spectral windows as an additional step
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 14 / 25
Proposed Algorithm
Additional step #2: Sparsifying and adding windowed ICAtraces
We propose sparsifying the spectra of the windowed ICAtraces by preserving the lobes around the top R peaks and
then adding these spectra which potentially can magnify thepeak corresponding to the pulse rate and suppress noisy
peaks.
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 15 / 25
Proposed Algorithm
Advantage of sparsifying and adding spectra illustrated
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 16 / 25
Proposed Algorithm
Advantage of sparsifying and adding spectra illustrated
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 16 / 25
Proposed Algorithm
Additional step #3: Constructing candidate PR trajectories
Candidate pulse rate trajectories are constructed byselecting R highest peaks from the sum of the sparsified ICAspectra. The trajectory whose cumulative spectral power is
the highest is declared as the best pulse rate trajectory
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 17 / 25
Proposed Algorithm
Illustration of constructing pulse rate trajectories (R=5)
1 2 3 4 5 6 7 8 9 10
0.8
1
1.2
1.4
1.6
1.8
2
2.2
analysis window
frequency (
in H
z)
Constructing frequency trajectories
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 18 / 25
Proposed Algorithm
Illustration of constructing pulse rate trajectories (R=5)
1 2 3 4 5 6 7 8 9 10
0.8
1
1.2
1.4
1.6
1.8
2
2.2
analysis window
frequency (
in H
z)
Constructing frequency trajectories
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 18 / 25
Proposed Algorithm
Illustration of constructing pulse rate trajectories (R=5)
1 2 3 4 5 6 7 8 9 10
0.8
1
1.2
1.4
1.6
1.8
2
2.2
analysis window
frequency (
in H
z)
Constructing frequency trajectories
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 18 / 25
Proposed Algorithm
Illustration of constructing pulse rate trajectories (R=5)
1 2 3 4 5 6 7 8 9 10
0.8
1
1.2
1.4
1.6
1.8
2
2.2
analysis window
frequency (
in H
z)
Constructing frequency trajectories
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 18 / 25
Proposed Algorithm
Illustration of constructing pulse rate trajectories (R=5)
1 2 3 4 5 6 7 8 9 10
0.8
1
1.2
1.4
1.6
1.8
2
2.2
analysis window
frequency (
in H
z)
Constructing frequency trajectories
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 18 / 25
Proposed Algorithm
Illustration of constructing pulse rate trajectories (R=5)
1 2 3 4 5 6 7 8 9 10
0.8
1
1.2
1.4
1.6
1.8
2
2.2
analysis window
frequency (
in H
z)
Constructing frequency trajectories
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 18 / 25
Proposed Algorithm
Illustration of constructing pulse rate trajectories (R=5)
1 2 3 4 5 6 7 8 9 10
0.8
1
1.2
1.4
1.6
1.8
2
2.2
analysis window
frequency (
in H
z)
Constructing frequency trajectories
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 18 / 25
Proposed Algorithm
Illustration of constructing pulse rate trajectories (R=5)
1 2 3 4 5 6 7 8 9 10
0.8
1
1.2
1.4
1.6
1.8
2
2.2
analysis window
frequency (
in H
z)
Constructing frequency trajectories
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 18 / 25
Proposed Algorithm
Illustration of constructing pulse rate trajectories (R=5)
1 2 3 4 5 6 7 8 9 10
0.8
1
1.2
1.4
1.6
1.8
2
2.2
analysis window
frequency (
in H
z)
Constructing frequency trajectories
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 18 / 25
Proposed Algorithm
Illustration of constructing pulse rate trajectories (R=5)
1 2 3 4 5 6 7 8 9 10
0.8
1
1.2
1.4
1.6
1.8
2
2.2
analysis window
frequency (
in H
z)
Constructing frequency trajectories
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 18 / 25
Proposed Algorithm
Illustration of constructing pulse rate trajectories (R=5)
1 2 3 4 5 6 7 8 9 10
0.8
1
1.2
1.4
1.6
1.8
2
2.2
analysis window
frequency (
in H
z)
Frequency trajectories
Chosen trajectory
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 18 / 25
Experiments and Results
Outline
1 Motivation
2 Data collection
3 Proposed Algorithm
4 Experiments and Results
5 Conclusions
6 Future work
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 19 / 25
Experiments and Results
Experimental setup I
T=20 sec and ∆T=1 sec was chosen. Rectangular, Hann, Hammingand Blackman windows were used. For sparsification, R=3,5,10,15were chosen.
FFTs of length 4096 were computed in finding the power spectra.
We consider the median of the values of PRW ,R(K ) for the given Rsand W s as the final decision on the PR, henceforth denoted asC SSPT
The code for the whole algorithm was written in MATLAB R2014a.
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 20 / 25
Experiments and Results
Experimental setup II
Root Mean Square Errors (RMSE) from the groundtruths were usedas the measurers of the performance of the proposed algorithm.
As a baseline scheme for Comparison, the work by Poh et al 2 is used.
Also to compare how sparsification can help, the spectra of the ICAtraces were added and then trajectories were constructed by pickingtop R peaks.
2Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard, “Non-contact, automatedcardiac pulse measurements using video imaging and blind source separation”, Optics express,vol. 18, no. 10, pp. 1076210774, 2010.
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 21 / 25
Experiments and Results
Results I
Mean RMSE in BPMRectangular Hann Hamming Blackman
7.91 9.49 8.88 10.03
Rectangular window gives the best performance (byconsidering R=3,5,10,15 and taking C SSPT) .
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 22 / 25
Experiments and Results
Results II
Mean RMSE in BPMR Sparse Spectra Full Spectra
3 8.30 8.77
5 8.12 8.36
10 8.18 8.43
15 8.15 8.21
It can’t be said with surety which R ensures bestperformance.
Benefit of sparsification is highlighted.
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 23 / 25
Experiments and Results
Results III
Mean (std. dev.) RMSE in BPMExperi- Baseline C SSPTmental Sparse Full
Condition spectra spectra
S3 0.5 11.36 (11.04) 6.45 (6.13) 7.88 (7.17)
S3 1 10.62 (7.77) 7.56 (8.04) 9.03 (8.68)
S3 2 19.08 (17.31) 13.24 (18.41) 13.06 (18.08)
3GS 0.5 13.75 (14.04) 6.98 (7.15) 7.79 (7.93)
3GS 1 15.57 (9.07) 7.86 (9.06) 7.82 (8.74)
3GS 2 17.38 (14.58) 15.41 (15.23) 15.25 (17.29)
The proposed SSPT framework outperforms the baseline framework by∼ 6.71 BPM on an average (by considering rectangular window and
R=3,5,10,15 for C SSPT).
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 24 / 25
Experiments and Results
Results IV
60
70
80
90
S U B J E C T 3
60
70
80
90
40
60
80
40
60
80
70
80
90
Hea
rtbea
t rat
e in
bea
ts p
er m
inut
e
10 20 30 4070
80
90
40
50
60
70
S U B J E C T 9
40
50
60
70
40
60
80
40
60
80
406080
100120140
10 20 30 40406080
100120140
Time (in second)
406080
100120140
S U B J E C T 11
S3_0.5
406080
100120140
3GS_0.5
50
100
150
S3_1
50
100
150
3GS_1
100
200
S3_2
10 20 30 40
100
200
3GS_2
Ground Truth Sparse Spectra Full Spectra Baseline
The benefit of the proposed method vs baseline method highlighted
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 25 / 25
Conclusions
Outline
1 Motivation
2 Data collection
3 Proposed Algorithm
4 Experiments and Results
5 Conclusions
6 Future work
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 26 / 25
Conclusions
Conclusions
Pulse rate estimation from face videos benefit from sparsifying thespectra of ICA traces and by using the fact that pulse rate is a slowlyvarying quantity.
The nearer the subject is to the camera, better is the estimationaccuracy.
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 27 / 25
Future work
Outline
1 Motivation
2 Data collection
3 Proposed Algorithm
4 Experiments and Results
5 Conclusions
6 Future work
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 28 / 25
Future work
Case for better trajectory selection
0 5 10 150
20
40
60
80
Trajectory index
RM
SE
of
ea
ch
tra
jecto
ry in
BP
M
RMSEs of all trajectories
← Least RMSE=0.70092BPM
← RMSE of chosen tra jectory=9.0377BPM
0 5 10 150.5
1
1.5
2
2.5
3
3.5x 10
4 Tρ values for all trajectories
Trajectory index
Tρvalue
← Tρ for least RMSE trajectory
← Tρ for selected tra jectory
A case for better a trajectory selection mechanism (where Tp is the
net spectral power of a trajectory across analysis windows)
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 29 / 25
Future work
Future work
Studying effects of using different video compression schemes(codecs) is also a part of our study in this direction.
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 30 / 25
Future work
Thank You!
Aditya Gaonkar P, Bhuthesh R, Dipanjan Gope, Prasanta Kumar Ghosh (Indian Institute of Science)Robust Real-Time Pulse Rate Estimation From Facial Video Using Sparse Spectral Peak TrackingSPCOM, 2016 31 / 25