Impact of Body Position on Imaging Ballistocardiographic Signals

Alexander Woyczyk

1 a

and Sebastian Zaunseder

1,2 b

Faculty of Information Technology, University of Applied Sciences and Arts Dortmund, Dortmund, Germany

Professorship for Diagnostic Sensing, Faculty of Applied Computer Science, University Augsburg, Augsburg, Germany

Keywords:

Imaging Ballistocardiography, Camera Based Monitoring, Physiological Signal Sensing, Heart Rate.

Abstract:

Current works direct at the unobtrusive acquisition of vital parameters from videos. The most common ap-

proach exploits subtle color variations. The analysis of cardiovascular induced motion from videos (imaging

ballistocardiography, iBCG) is another approach that can supplement the analysis of color changes. The

presented study systematically investigates the impact of body position (supine vs. upright) on iBCG. Our

research directs at heart rate estimation by iBCG and on the possibility to analyse ballistocardiographic wave-

forms from iBCG. We use own data from 30 healthy volunteers, who went through repeated orthostatic ma-

neuvers on a tilt table. Processing is done according to common procedures for iBCG processing including

feature tracking, dimensionality reduction and bandpass ﬁltering. Our results indicate that heart rate estima-

tion works well in supine position (root mean square error of heart rate estimation 5.68 beats per minute). The

performance drastically degrades in upright (standing) position (root mean square error of heart rate estima-

tion 21.20 beats per minute). With respect to analysis of beat waveforms, we found large intra-subject and

inter-subject variations. Only in few cases, the resulting waveform closely resembles the ideal ballistocardio-

graphic waveform. Our investigation indicates that the actual position has a large effect on iBCG and should

be considered in algorithmic developments and testing.

1 INTRODUCTION

Within the last decade, the processing of videos for

non-contact acquisition of vital parameters has de-

veloped into a large ﬁeld of research. Particularly

for cardiovascular and respiratory monitoring, the

technique offers far-reaching opportunities (Molinaro

et al., 2022; Zaunseder et al., 2018; Zaunseder and

Rasche, 2022). Researchers were able to acquire

respiration (van Gastel et al., 2016), oxygen satura-

tion (Moc¸o and Verkruysse, 2021), blood pressure

(BP) (Steinman et al., 2021) and local perfusion dy-

namics (Rasche et al., 2020) from videos. The vast

majority of available works directs at the estimation

of heart rate (HR) or heart rate variability (HRV), re-

spectively. The most common approach to derive HR

or HRV exploits subtle variations in the pixels’ in-

tensity. Such variations reﬂect the varying light ab-

sorption due to ﬁlling of superﬁcial blood vessels.

The similarity to the clinical photoplethysmography

(PPG) led to the name imaging photoplethysmogra-

phy (iPPG).

https://orcid.org/0000-0001-6026-8816

https://orcid.org/0000-0001-6114-3142

Instead of exploiting intensity variations, Balakr-

ishnan et al. have shown that HR can be determined

by tracking the location of feature points, i.e. not us-

ing intensity changes but exploit macro motion (Bal-

akrishnan et al., 2013). The technique closely relates

to the well known ballistocardiography (BCG) but

uses cameras (in reference to iPPG, we will name it

imaging ballistocardiography (iBCG) from here on).

Besides information on HR, BCG, particularly the

shape of the signal, carries information on other car-

diovascular quantities as cardiac output or ventricular

ejection time, rendering the technique very interest-

ing for cardiovascular monitoring (Inan et al., 2015;

Sadek, 2018). With respect to iBCG, few attention

was spent on a deeper characterization including the

factors inﬂuencing iBCG signals and on the possibil-

ities of using iBCG beyond HR.

Our work aims at a deeper characterization of

iBCG. Such investigation is important towards a po-

tential extended usage of iBCG, let it be alone or in

combination with iPPG. This paper presents prelim-

inary investigations on the general usability of iBCG

for HR and morphological analyses using iBCG in de-

pendency of the body position.

The remainder of the work is structured as fol-

Woyczyk, A. and Zaunseder, S.

Impact of Body Position on Imaging Ballistocardiographic Signals.

DOI: 10.5220/0011659900003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 4: BIOSIGNALS, pages 55-65

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

lows. First, we provide the background on BCG in

general and present processing approaches for iBCG

in particular. In section 3 we detail the pursued pro-

cessing method and evaluation metrics. In section 4

we present results on the HR estimation by iBCG and

analyse the resulting iBCG morphologies. Finally,

section 5 discusses our ﬁndings and relates them to

the literature.

2 BACKGROUND ON

BALLISTOCARDIOGRAPHY

2.1 Physiological Background

BCG describes minor movements of the body caused

by the blood ejection of the heart. It is closely con-

nected to central haemodynamics, as it measures the

effect of the force accelerating the blood from the

heart into the aorta and the subsequent directional

changes of the blood ﬂow. The terminology of the

ballistocardiographic waves was introduced by Starr

et al. and describe the body displacement along the

longitudinal, i.e. head-to-toe axis (Starr, 1958).

Among several small displacements, a healthy

BCG features a prominent movement complex (Fig-

ure 1, points I, J and K), which can be directly as-

sociated with the ejection of blood from the heart

and the following redirection of the blood’s move-

ment as it passes the aortic arch. The I-wave is

the effect of the force opposing the head-ward ac-

celeration of the blood being ejected from the left

ventricle into the aorta. Soon, the main portion of

the blood gets redirected and therefore accelerated in

foot-ward direction by passing through the aortic arch

and entering the abdominal aorta. The deceleration of

the head-ward and redirection to toe-ward movement

causes the J-wave. Lastly the K-wave is supposed to

originate from the deceleration of the blood stream

through the descending aorta. (Pinheiro et al., 2010)

Since the BCG directly originates from the blood

ejection, the acceleration and redirection of blood

ﬂow, it is of high interest to cardiovascular monitor-

ing. The BCG can be used to evaluate various phys-

iological and cardiac parameters, e.g. cardiac output

or BP (Javaid et al., 2016; Su et al., 2019). Because

of its periodicity with each heart beat it is possible to

measure HR and HRV. Since the BCG waves can be

linked to different stages of the cardiovascular cycle,

the analysis of the signals’ shape can yield further in-

formation, e.g. it can serve as an indicator of heart

disease or stenosis within the aorta (Pinheiro et al.,

2010). In combination with other physiological sig-

nals as PPG, the BCG can be used to determine pulse

-2

axial acceleration

(in mm/s

)

time in s

1.5

0.5

ECG

(in mV)

Figure 1: Illustration of the acceleration BCG (axial accel-

eration) signal together with the ECG. The BCG waveform

was modiﬁed from Pinheiro et al. (Pinheiro et al., 2010).

transit time (PTT). Examining the PTT can lead to

insight on vascular stiffness and is also connected to

BP (Leit

ao et al., 2018; Pielmus et al., 2021).

2.2 Conventional Hardware for

Ballistocardiography

Since early descriptions in the 1940s, there have been

different approaches to record BCG signals. The most

common setup is described by Starr and measures the

BCG along the longitudinal axis of the body. This is

done either standing on modiﬁed scales (Inan et al.,

2009) or sitting on a chair with piezoelectric sen-

sors (Alamets

a et al., 2004). Another common set-

up involves the subjects to lay in weight-sensitive

beds. As the force is recorded along the sagittal

axis (Soames and Atha, 1982), the setup results in a

different BCG morphology. Today, the possibilities

of small scale acceleration sensors and gyroscopes

present in ﬁtness trackers and smart watches allows

to measure BCG under various circumstances (Zhang

et al., 2021; Shin et al., 2022). The usage of videos

is yet another possibility to carry out BCG, which has

gained importance in recent years.

2.3 Imaging Ballistocardiography

A fundamental work on BCG measurement from

video sequences was presented by Balakrishnan et

al. (Balakrishnan et al., 2013). The core idea is to

use prominent features in the face area to estimate ax-

ial displacement caused by the ejection of blood from

the heart. Prominent features are typically pixels with

high image gradient as corners or edges and are iden-

tiﬁed in the ﬁrst frame. By using the Kanade-Lucas-

Tomasi (KLT) algorithm each feature’s new location

can be retrieved in the consecutive frames (Tomasi

and Kanade, 1991). The vertical motion trajectories

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

of several feature points are then used as input sources

for principal component analysis (PCA). In order to

determine the heart rate, Balakrishnan et al. select the

component with the highest periodicity among the de-

composed signals (Balakrishnan et al., 2013).

Shao et al. use a similar set-up, capturing the

face with a camera (Shao et al., 2017). Instead of

automatically searching for feature points, they man-

ually label prominent characteristics as moles, hair

or skin pigmentation and identify corner features us-

ing ”Good features to track” (Jianbo Shi and Tomasi,

1994). Those features are tracked using KLT and the

vertical trajectories are averaged over all features.

Hassan et al. also use KLT to track points within

the face (Hassan et al., 2017). Their selection of face

features consists of 32 points, horizontally aligned on

the forehead and another set of 16 feature points ver-

tically aligned on the nose ridge. The initial positions

are determined by using coordinates relative to a face

bounding box originating from the Viola-Jones face

detector (Viola and Jones, 2001). Using PCA they

deconstruct the multivariate signal into its main com-

ponents and evaluate the distribution of the main fre-

quency of each component in order to estimate the

HR.

Li et al. instead mount the camera onto the head

of their subjects (Li et al., 2020). Being able to use

the whole ﬁeld of view (FOV) there is a higher prob-

ability of detecting high-quality corner features for

tracking. Similar to Balakrishnan et al. (Balakrish-

nan et al., 2013) they use the KLT tracking algo-

rithm to determine movement of the head by the dis-

placement of prominent features in the subjects FOV.

Using PCA and frequency analysis they estimate the

heart rate as the frequency with the largest amplitude

within the range of 0.75 to 3 Hz.

Lee et al. (Lee et al., 2021) present a fusion ap-

proach of iPPG and iBCG to estimate HR. They use a

deep learning approach, i.e. single shot detector with

ResNet, to detect the face and track 80 points in a re-

duced facial area using the KLT tracker. As before,

the iBCG is then extracted using PCA. Finally, Lee et

al. combine iBCG information with iPPG to yield a

stable HR.

Taken together, the existing approaches except Li

et al. (Li et al., 2020) are closely related. Minor dif-

ferences regarding the area to be considered should

not have a strong impact as commonly facial features,

excluding the eyes, are used. The feature points itself

are typically deﬁned automatically according to the

detection of edges in the image. The further process-

ing is always based on the axial displacement. Dif-

ferences exist regarding additional processing steps,

particularly the usage of PCA or not.

Lastly the presented works differ in their ap-

proaches on estimating the HR from the calculated

iBCG signal, e.g. using the PCA-component with

the clearest main frequency (Balakrishnan et al.,

2013), peak frequency of the Fourier-transformed sig-

nal (Shao et al., 2017; Li et al., 2020; Lee et al., 2021)

or the mean of a normal distribution ﬁtted to a set of

observed peak frequencies (Hassan et al., 2017).

3 MATERIAL AND METHODS

3.1 Dataset

Overview: The used data originates from own

multimodal experiments invoking healthy subjects.

Throughout the experiment, each participant executed

one or two cold stress tests, went through repeated

epochs of paced breathing and repeated orthostatic

maneuvers. During the experiment we recorded mul-

tiple vital signs and three RGB videos. All subjects

gave written consent. The study was approved by the

Ethics Committee at TU Dresden (EK 311082018).

Procedure: The subjects were asked to lay down on a

tilt-table, where they were connected to the measure-

ment systems. Over a duration of approximately 49

minutes the tilt-table was alternated between supine

and upright position every 7 minutes to provoke a car-

diovascular reaction. The position change took ap-

proximately 10 seconds and was executed with the

subject on the table. Each orthostatic maneuver marks

the beginning of a new epoch (supine or upright), re-

sulting in a total of seven epochs for each subject.

Between orthostatic maneuvers, i.e. within each 7

minute epoch, participants had resting phases and ex-

ecuted at least one cold stress test or paced breathing

sequence. During the cold stress test subjects had to

submerge their left hand into a basin of cold water

(1 – 4

◦

C) for a duration of 60 seconds. The paced

breathing exercise consisted of following a rhythm of

8 breath cycles per minute for 60 seconds.

Measurements: During the experiment, we recorded

RGB videos and multiple vital signs. Videos were

recorded with three RGB cameras (UI-3060CP-C-HR

Rev 2 by IDS Imaging Development Systems GmbH;

Obersulm, Germany). All cameras were ﬁxed on a ro-

bust frame joint with the tilt-table and therefore static

in their relative position to the subject even during

tilting, as can be seen in ﬁgure 2. For this work,

only Camera 2 is of particular interest. This camera

recorded the subject’s head at a distance of approx-

imately 40 cm. The recorded area covered the head

and a small portion of the shoulders. Videos were

captured at a color depth of 12 bit, a frame rate of

Impact of Body Position on Imaging Ballistocardiographic Signals

Figure 2: Experiment setup: Subject on tilt table in upright

position. Three cameras were attached to the table’s frame,

additionally ECG, PPG, respiration and BP were recorded.

25 Hz and a spatial resolution of 1280 px × 960 px.

The videos were saved in a proprietary format with

lossless compression. The recordings took place in

a controlled environment using indirect artiﬁcial il-

lumination (Walimex pro LED Sirius 160 Daylight

65W by WALSER GmbH & Co. KG; Gersthofen,

Germany).

Several physiological signals were recorded to

serve as reference, namely PPG and respiration

(MP36 by Biopac; Goleta, United States of America)

as well as electrocardiography (ECG) and BP mon-

itoring system (Finapres Nova by Finapres Medical

Systems; Enschede, Netherlands). The signals of par-

ticular relevance for this study were the electrocar-

diogram (Einthoven II) and a ﬁnger PPG signal. We

used the PPG signal to assess HR estimation based on

iBCG. We used the ECG signal, particularly the time

instants of QRS complexes, in order to analyse the

waveform from iBCG. Details regarding the handling

of reference signals are given in section 3.3.

Used Data: Overall, we use data of 30 healthy sub-

jects (11 female, 19 male; age 20 – 59). From each

subject we use excerpts from six epochs, three epochs

in supine and three epochs in upright position. Each

excerpt has a length of two minutes. During that time,

subjects were at rest and quiet. There were no further

behavioural advises, still the facial expressions might

vary but we would not expect marked movements.

3.2 Video Processing

As stated before, most works on iBCG share a com-

mon methodology, i.e. using image features within

the face and processing their longitudinal displace-

ment. Following, we present our applied methodol-

ogy, which closely relates to the core procedure and

processing steps of Balakrishnan et al. We also tested

some modiﬁcations but we will only brieﬂy discuss

them in section 5.

3.2.1 Tracking Procedure

As ﬁrst step to capture head movements, we detect fa-

cial landmarks. We use the Viola-Jones face detection

to get a estimate of the faces position and size (Vi-

ola and Jones, 2001). Landmarks are then identiﬁed

within the face bounding box and are used to further

reﬁne the region of interest (ROI) in which tracking

points are located. There are 68 landmarks cover-

ing the jaw line, eyes, nose and mouth features. The

landmark detection in use is DLib’s HOG based im-

plementation of Kazemis ”Ensemble of Regression

Trees” (Kazemi and Sullivan, 2014; Dlib C++ Li-

brary, 2020). Instead of searching the whole face for

landmarks, we deﬁne a ROI covering an area between

mouth and eyes, i.e. we use landmarks at the in-

ner eye corners, mouth and nose. Within this ROI

we detect up to 60 feature points using the ”Good

features to track” algorithm (Jianbo Shi and Tomasi,

1994). The landmarks and feature points are identi-

ﬁed in the ﬁrst frame of each epoch’s video segment.

In the consecutive frames we use the KLT tracking

algorithm to calculate the displacement of each fea-

ture points (Tomasi and Kanade, 1991). Figure 3 il-

lustrates the face bounding box, landmark-based ROI

and selected feature points by an example.

3.2.2 iBCG Generation

The iBCG signal is generated by evaluating the tra-

jectories of the recorded feature points. As commonly

done throughout the literature, we used only the lon-

gitudinal axis, i.e. y-component, of each feature point

for signal generation. The further iBCG formation

follows the ideas of Balakrishnan et al. (Balakrish-

nan et al., 2013). We discard feature points with large

displacement in between frames as it indicates either

an unstable feature or possible movement caused by

facial expression. A large displacement is deﬁned as

a displacement that is higher than the mode of max-

imum displacements of each feature. This process-

ing step is supposed to eliminate features which can

pollute the signal with high displacement values and

therefore decrease signal quality. The remaining tra-

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

jectories are ﬁltered to reduce signal components un-

related to the heart’s beating activity. We use a band-

pass ﬁlter with cutoff frequencies at 0.75 and 5 Hz to

cover the signal and its ﬁrst harmonics, which have

an inﬂuence on the signal’s morphology. The result-

ing signals serve as input for PCA. PCA decomposes

the input signals into its main components, i.e. com-

ponents responsible for the most variance in the sig-

nal. The resulting components are then sorted by their

contribution to the signals variance. Finally the most

probable candidate component for iBCG is selected

by choosing the component with the highest periodic-

ity among the ﬁrst ﬁve components.

3.2.3 iBCG Post Processing

The further processing of iBCG signals differs ac-

cording to the aim of the analysis, i.e. HR estimation

or iBCG morphology analysis.

With respect to HR estimation from iBCG, we in-

troduce an additional bandpass ﬁltering step. As done

in comparable works (Balakrishnan et al., 2013; Lee

et al., 2021), we employ a ﬁth-order Butterworth ﬁl-

ter with cutoff frequencies at 0.7 Hz and 2.5 Hz. The

ﬁltered signal is then transformed to the frequency

domain using fast fourier transform (FFT). The FFT

is applied to sliding windows of 10 s at a displace-

ment of 1 s. The location of the highest peak between

40 and 180 Hz in the spectrum of each window yields

the estimated HR. Due to the overlap of the sliding

windows we calculate one HR estimate per second.

Each estimate is compared to a reference HR, which

is derived from the temporally aligned window ap-

plied to the synchronized reference signal (see sec-

tion 3.3.1 for details).

With respect to beat morphology, as not to impact

the waveform, we do not apply additional bandpass

ﬁltering.

3.3 Evaluation

3.3.1 Performance of HR Estimation

We use accuracy, root mean square error (RMSE),

Bland Altman plots and correlation plots to compare

the iBCG approach to a reference HR. The reference

HR is derived from the PPG by the following proce-

dure. First, we detect all beats in the PPG excerpt by

a customized version of L

azaro et al. (L

azaro et al.,

2013). In accordance to the HR estimation by iBCG,

we then apply a 10 s sliding window with a step size

of 1 s to the detections and derive the reference HR

from the median inter-beat distance in each window.

We then calculate the accuracy of our HR estima-

tion from iBCG signals in reference to the PPG HR.

Figure 3: Exemplary illustration of detected feature points,

which are subsequently tracked. The image shows the

Viola-Jones bounding box (red square), the ROI deﬁned

based on selected relevant facial landmarks (inner red poly-

gon) and the feature points (green dots).

The accuracy is deﬁned as the percentage of the esti-

mated HR (further denoted HR

(Alg)

) within 5 bpm of

the reference HR (further denoted as HR

(Ref)

RMSE is deﬁned by

RMSE =

∑

n=1



(Alg)

− HR

(Ref)



(1)

where n denotes a single window (in our case 10 s).

Though evaluating overlapping windows, we con-

sider each estimate equally as HR estimation was

done independently of previously calculated HR es-

timates. All subjects share the same number of win-

dows for each two-minute excerpt thus yielding equal

weight for each subject.

3.3.2 Assessment of Beat Morphology

In accordance to the analysis of BCG signals, our

morphology analysis addresses the shape of iBCG

beats in the second order derivative, i.e. the accel-

eration iBCG signal. In order to analyse the shape,

we construct beat templates for each epoch by group-

ing 10 consecutive beats. To delineate single beats,

we employ the synchronized reference ECG. We ﬁrst

detect QRS complexes using a customized version

of Pan Tompkins QRS detection algorithm (Pan and

Tompkins, 1985). Based on the occurrence of QRS

complexes, we extract iBCG segments by cutting the

signal midway in between QRS complexes. The sep-

arated segments are then aligned at the QRS complex’

R-peak and a template is generated by averaging the

signals of the iBCG segments. We then further im-

prove the template by correcting small phase delays

(less than 10 ms) utilizing the cross-correlation be-

tween preliminary template and each segment.

Impact of Body Position on Imaging Ballistocardiographic Signals

The quantitative evaluation of the waveforms di-

rects at the feasibility of morphological analysis of

iBCG signals and considers three aspects:

1. Beat-to-beat correlation within groups of 10 con-

secutive beats: A high beat-to-beat correlation co-

efﬁcient indicates a stable signal and thus good

signal quality including a prominent iBCG wave.

2. Template-to-template correlation within each

epoch: High correlation coefﬁcients between tem-

plates indicate a reliable signal quality throughout

the epoch. This indicator could serve as decision

criterion for comparing morphological structures

between epochs.

3. Template-to-template correlation between neigh-

bouring epochs: Since epochs alternate between

supine and upright position, low correlation coef-

ﬁcients indicate substantial differences in supine

and upright BCG morphology.

All correlation coefﬁcients were calculated after ac-

counting for small temporal displacements by using

cross-correlation. In addition, for qualitative analyses

we consider exemplary waveforms with respect to the

expectation on a conventional BCG waveform.

3.4 Comparative Method

We also evaluated vertical displacement trajectories

without using PCA, as suggested by Shao et al. (Shao

et al., 2017). We used up to 60 feature points, identi-

ﬁed by the ”Good features to track” algorithm within

a reduced face bounding box, i.e. cropped horizon-

tally by 25 % in order to diminish the possibility

of background within the ROI. Finally we employ

the same bandpass ﬁlter (cutoff frequencies 0.75 and

5 Hz) and follow the same post processing steps as

described in section 3.2.

4 RESULTS

4.1 Heart Rate Estimation

Figure 4 shows the results of the HR estimation for

each epoch. The ﬁgure illustrates a pronounced effect

of body position. Similarly, Table 1 shows the RMSE

for all epochs and supports the observation of notable

quality differences between the upright and supine

position. More in detail, Figure 5 shows the accu-

racy per subject and epoch. The graphic illustrates

three ﬁndings: ﬁrstly, as seen before, supine position

seems to be highly advantageous regarding accurate

HR estimation. Secondly, even on an intra-subject

Table 1: Overall RMSE of estimated HR grouped by body

position (all values in bpm).

Method Position Epochs

PCA supine 5.95 4.94 6.16

upright 21.78 19.40 22.42

without supine 19.97 18.86 17.01

PCA upright 35.49 34.18 37.53

basis the accuracy shows relevant changes between

different epochs of the same position. And thirdly,

some subjects behave markedly worse in most or all

epochs.

4.2 Beat Morphology Analysis

Figure 6 shows the mean inter-beat correlation coef-

ﬁcient of single beats for each epoch and each sub-

ject. Evaluating the beat-to-beat correlation coefﬁ-

cient shows a higher value for supine epochs than

those in upright position. This indicates a more

stable beat morphology in supine position. Cre-

ating template beats by grouping several consecu-

tive beats together and then calculating the cross-

correlation of templates within one epoch also sup-

ports our ﬁnding supine position being highly beneﬁ-

cial for iBCG recordings. Figure 7 shows for within-

subject comparison mostly a stronger inter-template

cross-correlation in supine epochs than in upright po-

sition. The high correlation demonstrates that the cre-

ated templates are similar throughout the whole epoch

and therefore suitable candidates for advanced BCG

analysis. Also Figure 8 demonstrates a low correla-

tion between templates of neighbouring, i.e. supine

vs. upright, epochs. Thus revealing that there are

substantial differences in templates from supine and

upright recordings.

4.3 Comparative Method

Using the trajectories without further signal decom-

position, i.e. PCA, we observed worse HR esti-

mations, as shown in table 1. Also the templates

generated with this method are less stable, than the

ones produced according to Balakrishnan et al. Even

though the results are inferior, it can still be observed,

that templates generated from epochs in supine body

position tend to be more stable than in upright posi-

tion.

5 DISCUSSION

Our investigations show that HR estimation by iBCG

works well in supine position. Our results are slightly

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

40 80 120

-60

-40

-20

17.24 bpm

-21.47 bpm

-2.12 bpm

50 100

100

120

(a) Results on ﬁrst supine epoch

40 80 120

-60

-40

-20

23.10 bpm

-55.79 bpm

-16.35 bpm

50 100

100

120

(b) Results on ﬁrst upright epoch

40 80 120

-60

-40

-20

12.11 bpm

-14.89 bpm

-1.39 bpm

50 100

100

120

40 80 120

-60

-40

-20

24.32 bpm

-51.42 bpm

-13.55 bpm

50 100

100

120

(d) Results on second upright epoch

40 80 120

-60

-40

-20

16.59 bpm

-18.24 bpm

-0.82 bpm

50 100

100

120

(e) Results on third supine epoch

40 80 120

-60

-40

-20

23.37 bpm

-55.02 bpm

-15.82 bpm

50 100

100

120

(f) Results on third upright epoch

Figure 4: Bland Altman plots and illustration of correlation for all epochs with y as reference HR and ˆy as estimated HR. Each

point in the plots represents a single 10 s window, each color a subject.

worse than reported in the literature but at an compa-

rable level. E.g. for using iBCG only, Lee et al. report

an RMSE of 3.48 bpm, 5.71 bpm and 15.86 bpm for

resting conditions, facial expressions and movements,

respectively (correlation coefﬁcients are 0.927, 0.920

and 0.051) (Lee et al., 2021). In general, with re-

Impact of Body Position on Imaging Ballistocardiographic Signals

supine 1

upright 1

supine 2

upright 2

supine 3

upright 3

subject number

0.2

0.4

0.6

0.8

Figure 5: Accuracy matrix showing values for all subjects

and epochs. There are distinct differences between supine

and upright epochs in almost all subjects.

supine 1

upright 1

supine 2

upright 2

supine 3

upright 3

subject number

-1

-0.5

0.5

Figure 6: Mean intra-epoch beat correlation on all consid-

ered subjects and epochs. The matrix shows the averaged

correlation coefﬁcient (after accounting for minor displace-

ment by using cross-correlation) calculated in-between all

beats of an epoch.

supine 1

upright 1

supine 2

upright 2

supine 3

upright 3

subject number

-1

-0.5

0.5

Figure 7: Inter-template correlation of all considered sub-

jects and epochs. The matrix shows the correlation coef-

ﬁcient (after accounting for minor displacement by using

cross-correlation) between all templates of an epoch.

Due to a delayed start of ECG recording of Subject 1 -

supine 1 only one template could be generated, therefore

no correlation was calculated.

spect to the absolute performance of HR estimation,

we have to emphasize that our work did not foster

an intense optimization, but we adhere to previously

supine 1

upright 1

supine 2

upright 2

supine 3

upright 3

subject number

-1

-0.5

0.5

Figure 8: Inter-epoch template correlation between neigh-

bouring epochs. The matrix shows the correlation coef-

ﬁcient (after accounting for minor displacement by using

cross-correlation) between all templates of one epoch with

all templates of the neighbouring epochs.

published procedures. We thus believe that the results

from section 4 could be improved, but this is not our

main aim within this work. We consider the found

differences between supine and upright position to be

of greater importance. The body position obviously

has a strong impact on the results as the unsatisfy-

ing performance in upright position clearly demon-

strates. We did consider other methods and feature

points as well, e.g. the method proposed by Shao et

al. (Shao et al., 2017). Shao et al. directly exploited

axial movements without calculating the acceleration

signal. The method thus basically skips some steps of

Balakrishnan et al. (Balakrishnan et al., 2013). How-

ever, using movements directly, our HR estimation

was drastically worse.

With respect to the signal morphology, our ﬁnd-

ings indicate the iBCG to be troublesome regardless

of the body position. While in some cases, as dis-

played in Figure 10, we could yield a waveform that

resembles the well known BCG shape, these cases

were the exception rather than the rule. Instead, we

often found periodically varying but heavily distorted

waveforms and large morphological inter-subject and

intra-subject differences, exemplary shown in Fig-

ure 9. This observation coincides with exemplary sig-

nals from most other works, which do not exhibit dis-

tinct BCG characteristics. While a stable waveform

and high correlation to the HR make such segments

very likely to originate from movement induced by

the heart motion, the deformation of the iBCG wave-

form complicate the analysis of BCG speciﬁc features

as measuring timings or amplitudes of the BCG com-

plex.

Lastly, we were not able to reproduce the ﬁnd-

ings of Shao et al., who provide examples of sig-

nal shapes that closely match the conventional BCG

waveform (Shao et al., 2017). One potential reason

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

Time (s)

Acceleration (a.u.)

-2

-1

-3

-2

-1

-3

-2

-1

-3

-2

-1

-3

Figure 9: Consecutive templates of varying beats, (subject

11, supine 3). Templates and beats are centered at ECG R-

peak.

might be the size of the observed area: if the spatial

resolution is too small, detailed ballistocardiographic

information might get blurred or lost completely. An-

other reason could be the careful manual selection of

a small ROI with a prominent characteristic as em-

ployed by Shao et al. (Shao et al., 2017), which leads

to a higher and more distinct waveform.

Concerning posture variations, we did expect dif-

ferences and some degradation. Jung et al. describe

a distinct impact of posture on the BCG morphol-

ogy (Jung et al., 2020). Similarly, Shin et al. re-

port relevant variation in BCG morphology between

standing, sitting and supine positions (Shin et al.,

2022). Though both works direct at wrist BCG, fa-

cial iBCG signals can be assumed to be affected by

positional changes and varying support of the head as

well. Against such background, our ﬁndings regard-

ing the signals’ shape are reasonable, though the (neg-

ative) effect is more pronounced than we expected.

Again, even with respect to the signals’ shape, we

would not rule out that methodological developments

or different recording parameters, e.g. higher reso-

lution or lower face to camera distance, might im-

prove the results. In fact, as the timing information

from iBCG provides highly relevant information, fu-

ture works should try to enhance and develop methods

towards more reliable iBCG waveforms.

6 CONCLUSIONS

Our analyses show that the body position has a rel-

evant impact on the quality of both, heart rate es-

timation and morphological analysis in iBCG. Such

ﬁndings are valuable for future research and develop-

ment, e.g. they indicate that more robust iBCG algo-

rithms should be developed and tested under varying

Time (s)

Acceleration (a.u.)

-0.5 0 0.5

-5

-3

-0.5 0 0.5

-5

-3

Figure 10: Recognizable BCG morphology (subject 18,

supine 1). Templates and beats are centered at ECG R-peak.

positions. As iBCG carries partially redundant and

partially complementary information to iPPG, iBCG

opens up wide opportunities for sensor data fusion

concepts. Particularly the possibility of constructing

PPG and BCG signals from the same (camera-)source

allows further investigations, as the temporal relation-

ship between heart activity and superﬁcial blood pul-

sation or the possibility to reduce BCG based artefacts

from iPPG. We consider such research and develop-

ment as very interesting (though challenging) as it has

the potential to indirectly improve future applications

of iPPG as BP estimation or skin perfusion imaging.

ACKNOWLEDGEMENT

This work was funded by the Deutsche Forschungs-

gemeinschaft (DFG, German Research Foundation),

project 401786308.

REFERENCES

Alamets

a, J., V

arri, A., Koivuluoma, M., and Barna, L.

(2004). The potential of emﬁ sensors in heart activ-

ity monitoring.

Balakrishnan, G., Durand, F., and Guttag, J. (2013). Detect-

ing pulse from head motions in video. Proceedings of

the IEEE Computer Society Conference on Computer

Vision and Pattern Recognition, pages 3430–3437.

Dlib C++ Library (2020). Dlib 19.21 Documentation.

Hassan, M. A., Malik, A. S., Foﬁ, D., Saad, N. M., Ali,

Y. S., and Meriaudeau, F. (2017). Video-based heart-

beat rate measuring method using ballistocardiogra-

phy. IEEE Sensors Journal, 17(14):4544–4557.

Inan, O. T., Etemadi, M., Wiard, R. M., Giovangrandi, L.,

and Kovacs, G. T. A. (2009). Robust ballistocardio-

gram acquisition for home monitoring. Physiological

Measurement, 30(2):169–185.

Inan, O. T., Migeotte, P.-F., Park, K.-S., Etemadi, M.,

Tavakolian, K., Casanella, R., Zanetti, J., Tank, J.,

Funtova, I., Prisk, G. K., and Di Rienzo, M. (2015).

Impact of Body Position on Imaging Ballistocardiographic Signals

Ballistocardiography and Seismocardiography: A Re-

view of Recent Advances. IEEE Journal of Biomedi-

cal and Health Informatics, 19(4):1414–1427.

Javaid, A. Q., Ashouri, H., Tridandapani, S., and Inan, O. T.

(2016). Elucidating the hemodynamic origin of bal-

listocardiographic forces: Toward improved monitor-

ing of cardiovascular health at home. IEEE Journal

of Translational Engineering in Health and Medicine,

4:1–8.

Jianbo Shi and Tomasi (1994). Good features to track. In

Proceedings of IEEE Conference on Computer Vision

and Pattern Recognition CVPR-94, pages 593–600.

IEEE Comput. Soc. Press.

Jung, H., Kimball, J., Receveur, T., Agdeppa, E., and

Inan, O. T. (2020). Quantiﬁcation of Posture-Induced

Changes in Bed-Based Ballistocardiogram. Comput-

ing in Cardiology, 2020-Septe:2020–2023.

Kazemi, V. and Sullivan, J. (2014). One millisecond face

alignment with an ensemble of regression trees. Pro-

ceedings of the IEEE Computer Society Conference

on Computer Vision and Pattern Recognition, pages

1867–1874.

azaro, J., Gil, E., Vergara, J. M., and Laguna, P. (2013).

Pulse rate variability analysis for discrimination of

sleep-apnea-related decreases in the amplitude ﬂuctu-

ations of pulse photoplethysmographic signal in chil-

dren. IEEE journal of biomedical and health infor-

matics, 18(1):240–246.

Lee, H., Cho, A., and Whang, M. (2021). Fusion Method

to Estimate Heart Rate from Facial Videos Based on

RPPG and RBCG. Sensors, 21(20):6764.

Leit

ao, F., Moreira, E., Alves, F., Lourenc¸o, M., Azevedo,

O., Gaspar, J., and Rocha, L. A. (2018). High-

resolution seismocardiogram acquisition and analysis

system. Sensors, 18(10):3441.

Li, F., Zhao, Y., Kong, L., Dong, L., Liu, M., Hui, M., and

Liu, X. (2020). A camera-based ballistocardiogram

heart rate measurement method. Review of Scientiﬁc

Instruments, 91(5).

Moc¸o, A. and Verkruysse, W. (2021). Pulse oximetry based

on photoplethysmography imaging with red and green

light. Journal of Clinical Monitoring and Computing,

35(1):123–133.

Molinaro, N., Schena, E., Silvestri, S., Bonotti, F., Aguzzi,

D., Viola, E., Buccolini, F., and Massaroni, C. (2022).

Contactless Vital Signs Monitoring From Videos

Recorded With Digital Cameras: An Overview. Fron-

tiers in Physiology, 13:160.

Pan, J. and Tompkins, W. J. (1985). A real-time qrs de-

tection algorithm. IEEE transactions on biomedical

engineering, (3):230–236.

Pielmus, A.-G., M

uhlstef, J., Bresch, E., Glos, M., Jun-

gen, C., Mieke, S., Orglmeister, R., Schulze, A., Sten-

der, B., Voigt, V., and Zaunseder, S. (2021). Sur-

rogate based continuous noninvasive blood pressure

measurement. Biomedical Engineering / Biomedi-

zinische Technik, 66(3):231–245.

Pinheiro, E., Postolache, O., and Gir

ao, P. (2010). Theory

and Developments in an Unobtrusive Cardiovascular

System Representation: Ballistocardiography. The

Open Biomedical Engineering Journal, 4(1):201–216.

Rasche, S., Huhle, R., Junghans, E., de Abreu, M. G., Ling,

Y., Trumpp, A., and Zaunseder, S. (2020). Associ-

ation of remote imaging photoplethysmography and

cutaneous perfusion in volunteers. Scientiﬁc Reports,

10(1):16464.

Sadek, I. (2018). Ballistocardiogram Signal Processing: A

Literature Review. arXiv preprint arXiv:1807.00951.

Shao, D., Tsow, F., Liu, C., Yang, Y., and Tao, N. (2017).

Simultaneous Monitoring of Ballistocardiogram and

Photoplethysmogram Using a Camera. IEEE Transac-

tions on Biomedical Engineering, 64(5):1003–1010.

Shin, S., Mousavi, A., Lyle, S., Jang, E., Youseﬁan, P.,

Mukkamala, R., Jang, D. G., Kwon, U. K., Kim, Y. H.,

and Hahn, J. O. (2022). Posture-Dependent Variabil-

ity in Wrist Ballistocardiogram-Photoplethysmogram

Pulse Transit Time: Implication to Cuff-Less Blood

Pressure Tracking. IEEE Transactions on Biomedical

Engineering, 69(1):347–355.

Soames, R. W. and Atha, J. (1982). Three-dimensional

ballistocardiographic responses to changes of pos-

ture. Clinical Physics and Physiological Measure-

ment, 3(3):169–177.

Starr, I. (1958). The relation of the ballistocardiogram to

cardiac function. The American Journal of Cardiol-

ogy, 2(6):737–747.

Steinman, J., Barszczyk, A., Sun, H.-S., Lee, K., and Feng,

Z.-P. (2021). Smartphones and Video Cameras: Fu-

ture Methods for Blood Pressure Measurement. Fron-

tiers in Digital Health, 3.

Su, B. Y., Enayati, M., Ho, K. C., Skubic, M., Despins, L.,

Keller, J., Popescu, M., Guidoboni, G., and Rantz, M.

(2019). Monitoring the relative blood pressure using

a hydraulic bed sensor system. IEEE Transactions on

Biomedical Engineering, 66(3):740–748.

Tomasi, C. and Kanade, T. (1991). Detection and Tracking

of Point Features. Technical report, Carnegie Mellon

University, Pittsburgh, PA.

van Gastel, M., Stuijk, S., and de Haan, G. (2016). Robust

respiration detection from remote photoplethysmog-

raphy. Biomedical Optics Express, 7(12):4941.

Viola, P. and Jones, M. (2001). Rapid object detection us-

ing a boosted cascade of simple features. In Proceed-

ings of the 2001 IEEE Computer Society Conference

on Computer Vision and Pattern Recognition. CVPR

2001, volume 1, pages I–511–I–518. IEEE Comput.

Soc.

Zaunseder, S. and Rasche, S. (2022). Chapter 7 - Clin-

ical applications for imaging photoplethysmography.

In Wang, W. and Wang, X., editors, Contactless Vital

Signs Monitoring, pages 149–164. Academic Press.

Zaunseder, S., Trumpp, A., Wedekind, D., and Malberg,

H. (2018). Cardiovascular assessment by imag-

ing photoplethysmography-a review. Biomedizinische

Technik, 63(5):529–535.

Zhang, Y., Zhang, X., Cui, P., Li, S., and Tang, J. (2021).

Key Feature Selection and Model Analysis for Blood

Pressure Estimation from Electrocardiogram, Ballis-

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

tocardiogram and Photoplethysmogram. IEEE Ac-

cess, 9:54350–54359.

APPENDIX

Accuracy and correlation matrices of the processing

steps without PCA as described in section 3.4 are pre-

sented below.

supine 1

upright 1

supine 2

upright 2

supine 3

upright 3

subject number

0.2

0.4

0.6

0.8

Figure 11: Without PCA processing: Accuracy matrix

showing values for all subjects and epochs. There are dis-

tinct differences between supine and upright epochs in al-

most all subjects.

supine 1

upright 1

supine 2

upright 2

supine 3

upright 3

subject number

-1

-0.5

0.5

Figure 12: Without PCA processing: Mean intra-epoch

beat correlation on all considered subjects and epochs. The

matrix shows the averaged correlation coefﬁcient (after ac-

counting for minor displacement by using cross-correlation)

calculated in-between all beats of an epoch.

supine 1

upright 1

supine 2

upright 2

supine 3

upright 3

subject number

-1

-0.5

0.5

Figure 13: Without PCA processing: Inter-template corre-

lation of all considered subjects and epochs. The matrix

shows the correlation coefﬁcient (after accounting for minor

displacement by using cross-correlation) between all tem-

plates of an epoch.

supine 1

upright 1

supine 2

upright 2

supine 3

upright 3

subject number

-1

-0.5

0.5

Figure 14: Without PCA processing: Inter-epoch tem-

plate correlation between neighbouring epochs. The ma-

trix shows the correlation coefﬁcient (after accounting for

minor displacement by using cross-correlation) between all

templates of one epoch with all templates of the neighbour-

ing epochs.

Impact of Body Position on Imaging Ballistocardiographic Signals