Explaining the Ergonomic Assessment of Human Movement in

Industrial Contexts

Sara Santos

1 a

, Duarte Folgado

1 b

, Jo

ao Rodrigues

, Naﬁseh Mollaei

Carlos Fuj

and Hugo Gamboa

1,2 c

Associac¸

ao Fraunhofer Portugal Research, Rua Alfredo Allen 455/461, 4200-135 Porto, Portugal

Laborat

orio de Instrumentac¸

ao, Engenharia Biom

edica e F

ısica da Radiac¸

ao (LIBPhys-UNL), Departamento de F

ısica,

Faculdade de Ci

encias e Tecnologia, FCT, Universidade Nova de Lisboa, 2829-516 Caparica, Portugal

Volkswagen Autoeuropa, Quinta da Marquesa, 2954-024 Q.ta do Anjo, Portugal

Keywords:

Ergonomics, Industry, Musculoskeletal Disorders, Inertial Sensors, Motion Capture, Risk Score.

Abstract:

The repetitive nature of manufacturing processes is identiﬁed as a risk factor for the onset of musculoskeletal

disorders. For prevention, the operator’s exposure risk is measured through ergonomic risk scores which are

often associated with a workstation, ignoring the variability among operators. Moreover, the score values

hinder a comprehensive interpretation by occupational physicians. Observation methods require signiﬁcant

effort, preventing accurate and continuous evaluation. The conducted study developed a solution using inertial

sensors for automatic operator risk exposure in the manufacturing industry. Two experimental assessments

were conducted: laboratory validation, performed by 14 subjects, using an optical motion capture system as

a reference; and ﬁeld evaluation, with 6 participants, acquired on a real automotive assembly line, served as

the basis for an ergonomic risk evaluation study. Through the research, it was implemented an upper-body

motion tracking algorithm relying on inertial information, to estimate the angular orientation of anatomical

joints. An adjusted ergonomic risk score, based on direct measurements was developed allowing an ergonomic

evaluation which also has an explanation approach, based on the comprehensive analysis of the angular risk

factors. Direct measurements fasten the ergonomic feedback, consequently, the evaluation can be extended to

more operators, ultimately preventing work-related injuries.

1 INTRODUCTION

Work-related musculoskeletal disorders (WMSDs)

represent a signiﬁcant portion of work-related health

problems in the European Union, impacting employ-

ees from different working sectors (Irastorza et al.,

2010). According to the World Health Organization,

musculoskeletal conditions are the second largest

contributor to disability worldwide and they are pre-

dicted to rise as the global population ages (Luttmann

et al., 2003). During 2017-2018, the upper limb or

neck disorders accounted for approximately 42% of

WMSDs, and within these, elbow diseases are the

most prevalent (H. Seidel et al., 2019).

In some industry sectors, e.g. textile and auto-

motive, production processes are typically based on

https://orcid.org/0000-0002-2554-3648

https://orcid.org/0000-0002-8481-6079

https://orcid.org/0000-0002-4022-7424

the cooperation between humans and machines. Al-

though the work methods carried out by workers have

predeﬁned motions and actions, their repetitive nature

can increase the risk of musculoskeletal disorders de-

velopment, leading not only to absenteeism but also

early retirement and loss of productivity (Uva et al.,

2008; Varandas. et al., 2019).

On large industrial environments, there are still

some unsolved challenges which prevent a more ef-

fective ergonomic job analysis. During the work

method design, which comprises a set of predeﬁned

motions, manufacturing industries rely on ergonomic

assessment tools which measure the workers’ risk ex-

posure through a risk score value. Thus, for a given

work method, the global risk score is calculated tak-

ing into account all local scores associated with each

motion or action that composes the work cycle. How-

ever, those are based on an average worker, mean-

ing that they do not take into account the variability

among operators that may exist at the manufacturing

Santos, S., Folgado, D., Rodrigues, J., Mollaei, N., Fujão, C. and Gamboa, H.

Explaining the Ergonomic Assessment of Human Movement in Industrial Contexts.

DOI: 10.5220/0008953800790088

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 4: BIOSIGNALS, pages 79-88

ISBN: 978-989-758-398-8; ISSN: 2184-4305

plant population, such as anthropometric variations,

operator’s age and work experience. Additionally, er-

gonomic teams might still rely on observational meth-

ods, which involve dedicated personnel to observe or

video record operators at work for posterior analysis.

Due to the high workload involved in this process, it

becomes unfeasible to employ observational methods

across the complete manufacturing population. More-

over, the outcome of the ergonomic risk assessment

results in a number which quantiﬁes the associated

risk yet, when occupational doctors receive their pa-

tients and wish to access the history of the assigned

workstations and associated ergonomic risk, they only

have access to a score to describe the risk, which is in-

sufﬁcient for an adequate analysis to the contributing

risk factors that lead to a given global risk score.

Although robots are becoming more common in

manufacturing environments, operators are still es-

sential. However, the concept of an operator is under-

going a paradigm shift through the new generation of

operators coming entitled ”Operator 4.0”. These new

smart and skilled workers will have ”super-strength”

provided by exoskeletons, smarter decision capabil-

ities supported by artiﬁcial intelligence, and able to

age healthily at work supported by a set of wearable

body monitoring devices (Romero et al., 2016).

2 RELATED WORK

Wearable devices have attracted considerable atten-

tion to industrial environments. By using inertial mo-

tion capture systems, data can be collected, and sev-

eral parameters can be directly measured, e.g. posi-

tion and velocity of each body segment, postural an-

gles trends and gait parameters, making these funda-

mental for ergonomics studies (Caputo et al., 2019;

Wang. et al., 2019).

Several surveys have been published concern-

ing human motion tracking as in (Filippeschi et al.,

2017) and (Pereira et al., 2017). Among the studies

which used human motion tracking methods for er-

gonomic assessment in industrial concepts, (Battini

et al., 2014) used solely inertial sensors to perform a

full-body ergonomic evaluation based on several er-

gonomic worksheets. (Peppoloni et al., 2016) and

(Vignais et al., 2017) focused on the ergonomic as-

sessment of the upper body regions but combined the

inertial measurement units (IMUs) with other meth-

ods such as electromyography, goniometers and video

system. On the other hand, (Bauters et al., 2018) re-

lied on a video system to perform a full-body analysis

to deliver operators productivity indicators.

In general, the studies fail to provide an error es-

timate for their motion tracking system. This can be

achieved by matching the proposed tracking methods

against other validated motion capture technique con-

sidered as ground truth. Moreover, the overall results

of ergonomics worksheets are uniquely a single score

value, lacking a more comprehensive analysis of the

risk factors.

There are some limitations associated with the use

of inertial sensors. For instance, gyroscope’s mea-

surements are affected by drift over time due to the

integration of device’s defects and noise (Beavers,

2017). Furthermore, sensor fusion techniques often

require to tune empirically the most adequate ﬁlter

parameters. Hence, one solution is to use a comple-

mentary ﬁlter with few tuning parameters to mitigate

the impact of gyroscope’s drift. Video systems also

present some challenges, e.g. the recordings are lim-

ited to deﬁned camera angles.

This work comprises the development of a mo-

tion tracking framework for a direct and continuous

ergonomic risk assessment on industrial contexts. We

intended to encompass a cost-effective solution to es-

tablish quantitative direct measurements of posture

and movement using inertial information, from ac-

celerometers, gyroscopes and magnetometers, for the

upper-body. Those measurements will be able to con-

tinuously monitor operators individually producing

also more comprehensive reports with explanations,

concerning the most contributing factors for the cal-

culated risk scores. Herewith, our work presents three

major contributions: (1) an upper-body motion track-

ing algorithm relying only on inertial information,

to estimate the absolute or relative angular orienta-

tion of anatomical joints; (2) the development of an

adjusted ergonomic risk score, based on direct mea-

surements, and (3) an ergonomic risk explanation ap-

proach, based on the comprehensive analysis of the

angular risk factors. It is expected that in long-term

this solution will help in the prevention of upper limb

WMSDs arising from repetitive tasks.

3 PROPOSED METHOD

When planning the implementation of a system using

direct methods for an ergonomic assessment, there are

typically three design considerations: explainability,

invasiveness and scalability. Explainability relates to

the degree of information that a setup can report. In-

vasiveness is related to the operator’s discomfort level

and also to the impact on the operator’s performance

due to the setup. Scalability establishes how many

subjects can, simultaneously, use the setup, depend-

ing on invasiveness and cost. We propose a system

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

which allows extracting information at an interme-

diate level, i.e. calculating low-level metrics of er-

gonomic risk and not demanding a large number of

sensors. Thus, it is expected that the system has an

average level of scalability, explainability and inva-

siveness.

3.1 System Overview

The developed upper body motion tracker system is

a sequential algorithm designed to obtain the time-

dependent angular information of several anatomical

segments. Since the upper limbs and spine are re-

gions with a higher prevalence and incidence of work-

related musculoskeletal disorders, the upper body was

the main focus of this research. Therefore, four

anatomical segments were deﬁned:

• Arm segment, as the segment between shoulder

and elbow joint;

• Forearm segment, as the segment between elbow

and wrist joint;

• Hand segment, as the segment between wrist and

distal region of the third metacarpal;

• Torso segment, as the segment between the jugu-

lar notch and the xiphoid process of the sternum.

The motion tracker implementation pipeline is de-

picted in Figure 1.

Data acquisition is the ﬁrst stage of the process.

Four IMU devices were employed and each of them

was attached to one of the four considered segments,

collecting acceleration, angular velocity and the mag-

netic ﬁeld data.

Signal processing methodology comprises pre-

processing and orientation estimation. The ﬁrst was

divided into two main processes: temporal synchro-

nisation, where equal sampling frequency and tempo-

ral alignment was ensured between the four IMU de-

vices, and noise reduction, through the implementa-

tion of a ﬁrst-order low-pass Butterworth ﬁlter with a

cutoff frequency of 1 Hz, on accelerometer and mag-

netic ﬁeld data. In its turn, orientation estimation de-

scribes the applied sensor fusion method and the nec-

essary considerations to obtain the angular informa-

tion of one segment relative to another or relative to

an anatomical plane.

The considered model admits ﬂexion/extension,

abduction/adduction, for shoulder joint; ﬂex-

ion/extension and pronation/supination for the

elbows; ﬂexion/extension and ulnar/radial deviation

for the wrist. Finally, the model also allows for the

torso ﬂexion/extension and lateral ﬂexion/extension.

Consequently, the whole model admits 8 degrees-

of-freedom and considers human movements of the

upper limb and torso.

3.2 Inertial Signal Acquisition

In the context of this research, 9-DoF IMUs contain-

ing a triaxial accelerometer, gyroscope and magne-

tometer were employed. The IMU devices collect

data sampled at 100 Hz and they were placed at the

following regions: IMU 1, IMU 2 and IMU 3 were

positioned at the posterior side of the hand, forearm

and arm, respectively. Particularly, IMU 2 was placed

in the wrist area and IMU 3 was located in the elbow

region. IMU 4 was positioned in the thorax area. To

assure a common axis alignment, the local axes direc-

tion of each device must be known before attaching

the device to the subject. It was considered that the

Y-axis, of all devices, points up. Figure 2 illustrates

the inertial devices placement.

Absolute Orientation

Acceleration

Magnetic Field

Angular velocity

Pre-processing Orientation Estimation

Relative Orientation

Frontal

Sagittal

Raw Sensor

Forearm - Hand

Arm - Forearm

Noise reduction

Temporal

synchronization

Sensor

Fusion

Figure 1: Upper-body motion tracker system framework.

Explaining the Ergonomic Assessment of Human Movement in Industrial Contexts

Figure 2: Placement of the IMU devices: three units on the

right upper limb and one unit on the torso. The devices were

commonly aligned with Y-axis pointing up.

3.3 Orientation Estimation

In order to estimate the anatomical segments attitude,

the signals gathered from accelerometers, gyroscopes

and magnetometers was combined through a sensor

fusion method. An algebraic algorithm, followed by

a quaternion-based complementary ﬁlter (QCF), de-

rived from (Colton, 2007; Valenti et al., 2015), were

implemented.

3.3.1 Algebraic Method

Data from accelerometer and magnetometer sensors

were combined using an algebraic algorithm (Lerner,

2012). Throughout the algorithm implementation, the

information of these two vectors was combined deﬁn-

ing an orthogonal coordinate system with the basis

vectors, expressed as a 3×3 rotation matrix. After-

wards, the rotation matrix can be translated into a ref-

erence quaternion, which represents the orientation of

a segment relative to Earth Reference Frame, follow-

ing East-North-Up conﬁguration. However, the ref-

erence quaternion does not represent the ﬁnal orienta-

tion, once it only relies on accelerometers and magne-

tometers readings. Nonetheless, this quaternion was

presented as measurements to QCF to obtain the ﬁnal

estimated quaternion.

3.3.2 Quaternion-based Complementary Filter

The quaternion-based attitude method updates the es-

timated quaternion through gyroscope’s measurement

and rectiﬁes it based on a reference quaternion from

the accelerometer and magnetometer measurements.

Using a quaternion representation of gyroscope’s

data and combining it with a previous instant es-

timated attitude quaternion, through the Hamilton

product, results in an update quaternion which rep-

resents the device rotation.

For initialising the ﬁlter, the update quaternion is

set equal to the reference quaternion. This way, both

of them represent the same device orientation. Never-

theless, for every sensor reading interval, a rectiﬁca-

tion and calculation of the estimated quaternion take

place. Next, we used interpolation to the reference

and the update quaternions. A Spherical Linear Inter-

polation (SLERP) (Dam et al., 1998) allows to weight

between the two quaternions. Once the gyroscope is

very accurate in short intervals it is more weighted.

Nevertheless, to stabilise the unwanted sensor drift, a

minor amount of the interpolation is directed towards

accelerometer and magnetometer, which are sensors

more trustworthy in the long term.

The estimated orientation exhibits the QCF char-

acteristics which combine high-frequency measures

from gyroscope and low-frequency from accelerome-

ters and magnetometers to deliver reliable motion in-

formation.

3.4 Angular Trajectory Reconstruction

After determining the estimated segment quaternion,

it is possible to make assumptions on the angular mo-

tion. It is assumed that consecutive IMU devices,

placed on the upper limb segments, are aligned, i.e.,

have one local axis that has the same direction. Di-

rection vectors can be expressed through pure quater-

nions in Sensor Frame. Making use of the dot product

between two vectors, the angle between segments is

determined.

Angular information between two consecutive

segments is deﬁned as relative orientation. On the

other hand, the angle between a segment and an

anatomical plane is deﬁned as the absolute orienta-

tion. The anatomical planes were deﬁned using the

local axes of an inertial device placed on a subject’s

torso. Additionally, the IMU device placed on the

torso segment is relevant to estimate the torso ﬂexion

and lateral ﬂexion. The angle of these last movements

is accomplished by comparing the torso’s current state

with torso’s rest position.

4 RESULTS

Two experimental assessments were conducted: labo-

ratory validation and ﬁeld evaluation. The laboratory

tests enabled the creation of a movements’ dataset,

where the proposed technique and a computer vision

approach were compared with an optical motion cap-

ture system. The ﬁeld evaluation dataset was acquired

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

on a real automotive assembly line and served as the

basis for the ergonomic risk assessment study.

4.1 Laboratory Validation

A validation protocol was designed to assess the pro-

posed method performance. Thus, to measure the

tracking error, the Vicon optical-passive motion cap-

ture system was used as a reference, which has a re-

ported error lower than 2 mm (Merriaux et al., 2017).

The proposed framework is intended to be used

through long-term acquisitions, corresponding to the

operator’s working shifts. However, sensor’s may

present some change in response over time, which

will be an issue. Another layer of information, to

periodically correct this sensor drift, might be a so-

lution. Despite video processing is computationally

more expensive than inertial sensor processing, it can

be used during short iterations to reset the drift from

sensors. Therefore, a video collection on the valida-

tion protocol is introduced to test and characterise the

computer vison-based library OpenPose (Cao et al.,

2018). The OpenPose is an open-source markerless

technology for multi-person 2D pose detection, iden-

tifying in total 135 keypoints, on single image, using

convolution neural network.

The acquisition protocol was performed by 14

subjects, nine men and ﬁve women, with an average

age of 26 ± 3 years. It allowed measuring the angu-

lar error across all considered joints in a wide range

of different movements. The validation protocol was

composed of two main parts: one describes a static

movement evaluation and the other details a dynamic

evaluation. The concepts static and dynamic denote if

the subject was standing or walking while doing the

designated movements, respectively.

Subjects wore a motion capture setup composed

of four IMUs sampling at 100 Hz and optical markers

tracked by Vicon cameras at 100 Hz. The Vicon setup

was composed of ten cameras, measuring an acquisi-

tion area of 8x4 m, and two standard cameras ﬁlming

the whole exercise, which were also used as input for

the OpenPose algorithm. The optical markers’ posi-

tions followed Vicon’s Upper Limb Model Guide de-

scriptions (Vicon Motion Systems, 2007). Raw data

is composed of 2 recording hours.

Several actions were manually segmented, specif-

ically ﬂexion/extension, abduction/adduction/, lateral

ﬂexion, ulnar/radial deviation and anatomical posi-

tion for static evaluation; ﬂexion and anatomical posi-

tion for dynamic trials. Figure 3 exhibits an example

of an angular reconstruction, representing the perfor-

mance of QCF and OpenPose. It can be observed that

both methods reconstruct the motion similarly to Vi-

con however, it can be noted an offset from OpenPose

during abduction and from QCF during anatomical

position.

Time/ s

Angle/ °

Arm Abduction

Anatomical

Position

Action

Anatomical

Position

Abduction

Figure 3: Angular reconstruction of arm’s abduction and

adduction. QCF (green), Vicon (black) and OpenPose

(blue) results.

It is relevant to explain that the procedure adopted

to adjust the light conditions of the tests is complex.

On one hand, the best conditions for using Vicon re-

quire low ambient light, while on the other hand, the

conditions for using OpenPose require regular ambi-

ent light so that the subject’s skeletal image contours

can be identiﬁed by the model. It was decided to

minimise Vicon’s error, since it was considered the

ground truth of this study and low ambient light con-

ditions were applied. However, this fact degraded

the performance of OpenPose algorithm, and conse-

quently the hand segment had to be neglected due to

inadequate low light conditions.

To perform a quantitative performance assessment

of both methods, two evaluation metrics were used:

the Cumulative Distribution Function (CDF) and the

Root-Mean-Square Deviation (RMSE). The function

represented in equation (1) is the CDF of a real-valued

random variable X,

(x) = P(X ≤ x) (1)

where P(X ≤ x) is the probability that the considered

variable X takes on a value less than or equal to x.

Algorithms can be analysed using their RMSE as

a measure of how well they describe a given set of ob-

servations. Equation (2) represents the RMSE, where

denotes the groundtruth value at time t provided by

Vicon and ˆy

denotes the predicted value at time t es-

timated by the upper-body tracking method.

RMSE =

∑

t=1

− ˆy

)

(2)

Firstly, the CDFs were calculated to assess each

segment performance under the QCF and OpenPose

Explaining the Ergonomic Assessment of Human Movement in Industrial Contexts

ArmTorso

Forearm

QCF

QCFQCF

QCF

Figure 4: Cumulative distribution function for the absolute error of OpenPose and QCF across torso, arm and forearm seg-

ments. Top: static evaluation; Bottom: dynamic evaluation.

methods. Figure 4 represents the CDF for QCF and

OpenPose techniques. Through the analysis of the

CDF for the static tests, it is possible to conclude that

arm and forearm’s movements present a lower error

when assessed with OpenPose algorithm. However,

the torso’s reconstruction shows better results with

QCF.

Contrary to static tests, in dynamic trials Open-

Pose presents better results for torso movements and

QCF has a higher performance for arm and forearm

exercises.

Table 1 represents the RMSE results for both static

and dynamic trials.

Table 1: Root mean square error regarding QCF and Open-

Pose methods. Static and dynamic evaluations.

QCF OpenPose

RMSE (

◦

) RMSE (

◦

)

Static Dynamic Static Dynamic

Torso 22 21 25 23

Arm 18 27 13 29

Forearm 30 20 20 27

In general, the table allows inferring that QCF and

OpenPose have similar performance. The arm seg-

ment presents the lowest movement’s error in static

trials. In its turn, the forearm segment overall results

show a higher error in both algorithms when compar-

ing it Vicon’s reference.

4.2 Field Evaluation

This section of the study consisted of an ergonomic

risk assessment of three workstations at a real auto-

motive assembly line, comprised by repetitive work

cycles.

In this study, we recruited six participants, four

men and two women, with an average age and height

of 31 ± 8 years and 173 ± 6 cm, respectively, without

any known musculoskeletal pathology. The operators

were asked to perform their working tasks while using

IMUs attached to their body. Prior to the experiment,

subjects signed and obtained a participation informed

consent. This research reports the results from three

workstations, Liftgate, Fender and Doors, from the

Bodyshop assembly line. Operators wore four IMUs

and were asked to perform two calibration positions,

N-pose and T-pose, in the beginning, and at the end

of the test. The curated dataset is composed of 4.23

recording hours.

4.2.1 General Workstation Risk

Before adopting strategies to improve working condi-

tions, situations that can contribute to operators’ risk

must be identiﬁed. Ergonomic indexes grant infor-

mation on the main risk factors, allowing to prioritise

interventions. The Rapid Upper Limb Assessment

(RULA) worksheet can be used to screen and identify

harmful postures (McAtamney and Corlett, 1993).

In this research, we developed an adapted version

of RULA’s, named Adjusted Rapid Upper Limb As-

sessment (AdRULA), which was implemented. The

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

Score2

Score3

60%

Score4

35%

Score5

Liftgate

Score2

Score3

70%

Score4

26%

Score5

Fender

Score2

17%

Score3

68%

Score4

15%

Doors

Figure 5: Liftgate, Fender and Doors workstation analysis. Mean score distribution for each workstation.

Liftgate

Fender

Doors

200

150

-50

-100

100

200

150

-50

100

200

150

-50

-100

100

Forearm Arm Torso Hand

+1 +2 +1 +2 +3 +4 +1 +2 +3

AdRULA

Angle/ °Angle/ °Angle/ °

Figure 6: Representation of operator’s average ﬂexion and extension movements from Liftgate, Fender and Doors worksta-

tions with AdRULA score thresholds.

AdRULA focus on the subject’s upper-body selects

postures every 0.02 s and apprehends poses via direct

measurements, e.g., wearable technology. The local

and ﬁnal scores are determined similarly to the RULA

method.

The average workstation’s score, using AdRULA

index, summarised in Figure 5, was determined

through two steps: 1) considering a single operator

performing the actions of a selected workstation, in

each work cycle, it was determined the percentage of

time spent in each score and, afterwards the individual

average was obtained; 2) the mean score for a given

workstation is ﬁnally calculated by averaging all the

operators scores (calculated in the previous step).

The charts demonstrate that, in general, when op-

erators perform tasks in the considered workstations

they stand for a longer period in a level 3-4 risk

zone which represents a low risk. Despite being a

Explaining the Ergonomic Assessment of Human Movement in Industrial Contexts

small percentage, the Liftgate and Fender worksta-

tions present a level 5 risk. Accordingly, those work-

stations represent a higher risk to operators in terms

of postures. We anticipated these results since Lift-

gate and Fender are workstations which require pos-

tures more prone to a higher exposure risk, such as

overhead motions.

4.2.2 General Workstation Explanation

From an ergonomic perspective, it is relevant to iden-

tify which movements contribute to a higher risk of

injuries. For the rest of this section, we present a

more detailed analysis of the factors that contribute to

the global scores, in order to complement the analysis

with a more interpretable and explainable methodol-

ogy. Figure 6 represents the distribution of extension

and ﬂexion movements for each workstation, in the

form of a probability density of these data.

It can be observed that the torso’s movements have

similar angular distribution for the considered work-

stations. Moreover, for Liftgate and Fender, the hand

movements have a higher probability of performing

ﬂexion exercises around 50

◦

while in the Doors work-

station the highest probability stands in the 25

◦

range.

Flexions and extensions distribution between arm and

forearm segments present evident unlikeness. While

in the Liftgate, the forearm presents a highest den-

sity probability between 50

◦

and 100

◦

, Fender and

Doors present two prominent probability peaks: 25

◦

and 50

◦

for Fender; 25

◦

and 90

◦

for Doors. Overall, as

working conditions are more demanding for arms and

hands, the Liftgate is classiﬁed with higher scores.

Doors workstation, with a larger probability of pos-

tures around the segment’s neutral zone, is evaluated

with lower levels.

4.2.3 Team Explanation

While working in the same workstation, operators

might not share the same characteristics, e.g., height,

weight, limbs length, and others. Figure 7 represents

the probability density of four different subjects per-

forming the tasks assigned to the Liftgate workstation.

Subjects’ characteristics are also depicted. Through-

out the analysis, it can be reasoned that among op-

erators from the same workstation, which have dif-

ferent characteristics, angular movements distribution

is not identical. Consequently, the individual’s er-

gonomic risk will be different from the one that could

be assigned to an average worker. The score value

might not be simple to interpret and consequently,

hinder occupational doctors and team leaders to per-

ceive operators’ needs. The individual analysis helps

to understand if the operator performs tasks within the

workstation risk range or if their characteristics inten-

sify/mitigate the risk. Thus, having personal reports,

with detailed movements information, can be an ad-

vantage for improving injuries-preventive recommen-

dations and for adjusting work conditions.

5 CONCLUSIONS

WMSDs represent a signiﬁcant portion of work-

related health problems, affecting workers from all

sectors. This research provided three major contribu-

tions. Firstly, an upper-body human motion tracking

algorithm using inertial sensor information was used

to estimate the absolute and relative orientation of

anatomical joints. Secondly, an adjusted ergonomic

risk score was developed based on direct measure-

ments. Finally, an ergonomic risk explanation ap-

proach, based on the comprehensive analysis of the

angular risk factors was presented.

Several conclusions were established using the

validation dataset. The OpenPose approach was used

as a mocap method, with similar performance to QCF,

yet it has some challenges. The ﬁrst is that OpenPose

is a computationally expensive algorithm. Secondly,

it is also prone to error in the presence of occlusion

(when the algorithm fails to track a limb).

Employing the estimated orientation of anatomi-

cal joints, provided by the system, it is possible to

conduct an ergonomic risk assessment. The work-

stations that presented a higher level of risk, Lift-

gate and Fender, behold actions that, effectively, re-

quire positions more susceptible to risk, e.g. over-

head work. Nowadays, the global risk score is often

agnostic to the variability of operators’ characteris-

tics and the scores, assigned based on ergonomics as-

sessments, use as reference an average worker. While

completing the risk analysis, it is possible to point out

evident motion differences among operators who per-

form the same workstation’s tasks. Hence, an individ-

ual ergonomic approach is better suited for prevent-

ing injuries, once it can unmask risk factors exposure.

The evaluation should be individual-related and not

the collective. Whenever risk exposure management

is a concern, the ergonomic analysis should be avail-

able for each worker. At last, providing explainability

to risk assessments is an added value to occupational

doctors once it allows a more comprehensive analysis

which can be relevant to support the decision-making

process for different strategies that can be addressed

to the worker by the Team Leader and/or occupational

doctor.

Comparing to previous researches presented in lit-

erature (Battini et al., 2014; Vignais et al., 2017) this

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

Subject 1

Angle/ °

200

150

-50

-100

Gender: Female

Age: 33

Height: 169 cm

Arm length: 28 cm

Forearm length: 24 cm

100

Angle/ °

200

150

-50

-100

100

Gender: Male

Age: 29

Height: 169 cm

Arm length: 26 cm

Forearm length: 22 cm

Forearm Arm Torso Hand

Gender: Male

Age: 23

Height: 180 cm

Arm length: 36 cm

Forearm length: 26 cm

Gender: Male

Age: 27

Height: 180 cm

Arm length: 32 cm

Forearm length: 25 cm

Subject 2

Subject 3

Subject 4

Angle/ °

200

150

-50

-100

100

Angle/ °

200

150

-50

-100

100

+1 +2 +1 +2 +3 +4 +1 +2 +3

AdRULA

Figure 7: Comparison of average ﬂexion and extension movements distribution from four different subjects while performing

Liftgate’s tasks, with AdRULA score thresholds. Right side - subjects’ characteristics.

study provides a quantitative error estimate of the pro-

posed tracking system, accomplished through a lab-

oratory validation. We combined different analysis

tools regarding the two experimental assessments, the

laboratory validation and the ﬁeld evaluation. The tri-

als required to design and organise two protocols in

which sensors attachment, calibration and monitoring

were detailed and contained guidelines for both par-

ticipants and specialists. Furthermore, the provided

ergonomic study is not only concerned with deliver-

ing a global risk analysis of a workstation, but it also

reaches the individual level.

With the current work, we can conclude that the

proposed method is feasible in a real manufacturing

context and provides a faster ergonomic analysis. Ac-

cordingly, we encourage the use of inertial sensors

as an effective method for detailed ergonomic assess-

ment in industrial environments. Nevertheless, the

study has a limited sample size thus, to improve ro-

bustness, the data collection should be increased.

As future work, it is expected to reduce the sen-

sor fusion accumulated errors of the long term system

which arise from the need of sensors to re-calibrate.

A solution would be introducing multimodal sensor

Explaining the Ergonomic Assessment of Human Movement in Industrial Contexts

fusion approaches, e.g. using video recordings to pe-

riodically calibrate the system. Video recordings are

limited to deﬁned camera angles however, that shall

not be a problem since during repetitive tasks we can

anticipate the action places and position the cameras

accordingly. Additionally, the explanation approach

can be increased through the analysis of three dimen-

sions of risk factors exposure: intensity, duration and

frequency.

REFERENCES

Battini, D., Persona, A., and Sgarbossa, F. (2014). Inno-

vative real-time system to integrate ergonomic evalu-

ations into warehouse design and management. Com-

puters & Industrial Engineering, 77:1 – 10.

Bauters, K., Cottyn, J., Claeys, D., Slembrouck, M., Vee-

laert, P., and van Landeghem, H. (2018). Auto-

mated work cycle classiﬁcation and performance mea-

surement for manual work stations. Robotics and

Computer-Integrated Manufacturing, 51.

Beavers, I. (2017). The Case of the Misguided

Gyro. https://www.analog.com/en/analog-dialogue/

raqs/raq-issue-139.html. Accessed: 2019-12-05.

Cao, Z., Hidalgo, G., Simon, T., Wei, S.-E., and Sheikh,

Y. (2018). OpenPose: realtime multi-person 2D pose

estimation using Part Afﬁnity Fields. In arXiv preprint

arXiv:1812.08008.

Caputo, F., Greco, A., D‘Amato, E., Notaro, I., and Spada,

S. (2019). IMU-Based Motion Capture Wearable Sys-

tem for Ergonomic Assessment in Industrial Environ-

ment. Springer Nature, pages 215–225.

Colton, S. (2007). The Balance Filter - A Simple Solution

for Integrating Accelerometer and Gyroscope Mea-

surements for a Balancing Platform.

Dam, E. B., Koch, M., and Lillholm, M. (1998). Quater-

nions, interpolation and animation. Technical Re-

port DIKU-TR-98/5, Department of Computer Sci-

ence University of Copenhagen, Universitets parken

1DK-2100Kbh, Denmark.

Filippeschi, A., Schmitz, N., Miezal, M., Bleser, G., Ruf-

faldi, E., and Stricker, D. (2017). Survey of motion

tracking methods based on inertial sensors: A focus

on upper limb human motion. Sensors, 17.

H. Seidel, D., Ditchen, D., Hoehne-H

uckst

adt, U., Rieger,

M., and Steinhilber, B. (2019). Quantitative measures

of physical risk factors associated with work-related

musculoskeletal disorders of the elbow: A systematic

review. International Journal of Environmental Re-

search and Public Health, 16:130.

Irastorza, E. S., Xabier, and Copsey, S. (2010). OSH in

ﬁgures: Work-related musculoskeletal disorders in the

EU — Facts and ﬁgures. European Agency for Safety

and Health at Work.

Lerner, G. M. (2012). Three axis attitude determination.

In Wertz, J., editor, Spacecraft Attitude Determination

and Control, volume 73 of Astrophysics and Space

Science Library, chapter 12, pages 420–435. Springer

Netherlands.

Luttmann, A., J

ager, M., Griefahn, B., Cafﬁer, G., and

Liebers, F. (2003). Preventing Musculoskeletal Dis-

orders in the Workplace. World Health Organization.

McAtamney, L. and Corlett, E. N. (1993). Rula: a sur-

vey method for the investigation of work-related up-

per limb disorders. Applied Ergonomics, 24(2):91 –

99.

Merriaux, P., Dupuis, Y., Boutteau, R., Vasseur, P., and Sa-

vatier, X. (2017). A study of vicon system positioning

performance. Sensors, 17(7).

Peppoloni, L., Filippeschi, A., Ruffaldi, E., and Avizzano,

C. (2016). A novel wearable system for the online

assessment of risk for biomechanical load in repeti-

tive efforts. International Journal of Industrial Er-

gonomics, 52:1 – 11. New Approaches and Interven-

tions to Prevent Work Related Musculoskeletal Disor-

ders.

Pereira, A., Guimar

aes, V., and Sousa, I. (2017). Joint an-

gles tracking for rehabilitation at home using inertial

sensors: a feasibility study. In Proceedings of the 11th

EAI International Conference on Pervasive Comput-

ing Technologies for Healthcare, pages 146–154.

Romero, D., Stahre, J., Wuest, T., Noran, O., Bernus, P.,

Fasth, Fast-Berglund, A., and Gorecky, D. (2016). To-

wards an Operator 4.0 Typology: A Human-Centric

Perspective on the Fourth Industrial Revolution Tech-

nologies. In International conference on computers

and industrial engineering (CIE46) proceedings.

Uva, A., Carnide, F., Serranheira, F., Miranda, L., and

Lopes, M. (2008). Les

oes m

usculoesquel

eticas rela-

cionadas com o trabalho-guia de orientac¸

ao para a

prevenc¸

ao. Direcc¸

ao Geral da Sa

ude.

Valenti, R., Dryanovski, I., and Xiao, J. (2015). Keeping

a good attitude: A quaternion-based orientation ﬁlter

for imus and margs. Sensors, 15:19302–19330.

Varandas., R., Folgado., D., and Gamboa., H. (2019). Eval-

uation of spatial-temporal anomalies in the analysis of

human movement. In Proceedings of the 12th Inter-

national Joint Conference on Biomedical Engineering

Systems and Technologies - Volume 4: BIOSIGNALS,,

pages 163–170. INSTICC, SciTePress.

Vicon Motion Systems (2007). Upper Limb Model. Product

Guide. http://www.vicon.com.

Vignais, N., Bernard, F., Touvenot, G., and Sagot, J. C.

(2017). Physical risk factors identiﬁcation based on

body sensor network combined to videotaping. Ap-

plied Ergonomics, 65.

Wang., H., Mohamed Refai, M., and van Beijnum., B. F.

(2019). Measuring upper-extremity use with one imu.

In Proceedings of the 12th International Joint Confer-

ence on Biomedical Engineering Systems and Tech-

nologies - Volume 4: BIOSIGNALS,, pages 93–100.

INSTICC, SciTePress.

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing