International Journal of Industrial Ergonomics 64 (2018) 122e133

Contents lists available at ScienceDirect

International Journal of Industrial Ergonomics

journal homepage: www.elsevier.com/locate/ergon

A design framework for the mass customisation of custom-fit bicycle

helmet models
Thierry Ellena a, *, Helmy Mustafa a, Aleksandar Subic b, Toh Yen Pang a
a
School of Aerospace, Mechanical and Manufacturing Engineering, RMIT University, Building 251, Level 3 Bundoora Campus East, Plenty Road, Bundoora
VIC 3083, Australia
b
Swinburne Research & Development, Swinburne University of Technology, Melbourne, Australia

a r t i c l e i n f o a b s t r a c t

Article history: Mass customisation (MC) can provide significant benefits to the customers. For example, custom-fit
Received 1 December 2016 design approaches can improve the users’ perceived comfort of products where the fit is an important
Received in revised form feature. MC can also bring major value to the producers, where for instance, premium prices can be
11 October 2017
implemented to the products. Research show that MC can bring competitive advantages especially when
Accepted 12 January 2018
Available online 2 February 2018
the system is new. It is therefore surprising that MC of helmets has not been studied more extensively,
especially given the advances in 3D scanning, computational analyses, parametric design, and additive
manufacturing techniques. The purpose of this study was to present a novel MC framework for the
Keywords:
Mass customisation
design of custom-fit bicycle helmet models.
Custom-fit In the proposed design framework, we first categorized a subset of the Australian population into four
Helmet groups of individuals based on their similar head shapes. New customers were then classified inside one
Product design of these groups. The customisation took place inside these groups to ensure that only small variations of
Finite elements analysis the helmet liner were implemented. During the design process, the inside surfaces of a generic helmet
model was modified to match the customer's head shape. We demonstrated that all the customized
models created complied with the relevant drop impact test standard if their liner thickness was be-
tween the worst and best case helmets of each group. Fit accuracy was verified using an objective
evaluation method. Future work should include detailed description of the manufacturing methods
engaged in our MC framework.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction As summarised by Da Silveira and co-authors in the well

documented MC reviews in (Da Silveira et al., 2001) and (Fogliatto
Mass customisation (MC) aims at providing customised prod- et al., 2012), multi degrees of mass customisation exist, from full
ucts or services to consumers in large volumes and at costs customer product specifications, to simple options selection.
reasonably low compared to conventional customisation processes Although a fully individualisation (e.g. bespoke tailoring) can hold
(Davis, 1989; Pine, 1993; Da Silveira et al., 2001). More specifically, more value to the consumer, often compromises must be reached
MC systems seek to reach customers as in the mass produce market to an acceptable level of customisation for a specific product.
but try to consider them individually as in the on-one production Therefore, MC should be a good mix between standardisation and
method. The reasoning behind the growth of MC systems in the late individualisation to be successful (Westbrook and Williamson,
1980s, early 1990s was threefold; (i) the development of advanced 1993).
manufacturing technologies, (ii) the increased demand for diversity Multiple generic levels of mass customisation have been pro-
in the products’ range and, (iii) the collapse of many mass in- posed in the past (Pine, 1993; Lampel and Mintzberg, 1996; Gilmore
dustries (Pine, 1993; Hart, 1995; Kotha, 1995), all increasing the and Pine, 1997; Spira, 1993). Da Silveira et al (Da Silveira et al.,
need of production methods focusing further on the individual. 2001). condensed the proposed classifications in an eight-level
MC scheme, ranging from whole customisation (consumer create
the product in collaboration with the designer) to whole stand-
ardisation. MC products between these two extreme categories
* Corresponding author.
E-mail address: [email protected] (T. Ellena). might be individualised at either the fabrication level (customer-

https://doi.org/10.1016/j.ergon.2018.01.005
0169-8141/© 2018 Elsevier B.V. All rights reserved.
 T. Ellena et al. / International Journal of Industrial Ergonomics 64 (2018) 122e133 123

tailored products), the assembly level (modular components), the Merle et al. (2008, 2010). investigated further the works from
delivery level (simple addition), the distribution level (different (Addis and Holbrook, 2001; Squire et al., 2004) to identify the
packaging), or the usage level (customers can alter the product drivers of MC value from the consumers' perspectives. They
during use). Evaluating the appropriate level of individualisation demonstrated that the perceived value increases as a result of the
for a specific product can be difficult. Ideally, preliminary studies intrinsic and extrinsic characteristics of customised products. On
should assess the customers’ interest to a level of customisation, the one hand, the utilitarian value (obtaining a product which
measure the feasibility to deliver at this level, and determine if matches one's preferences the closest), the uniqueness value (dis-
achieving such level holds comparative advantages. tinguishing oneself from others via the mass-customised product)
In this paper, we present a MC framework for the design of and the self-expressiveness value (obtaining a product that repre-
custom-fit bicycle helmets. Custom-fit meaning personalised in sents oneself) are great benefits conveyed by MC systems. On the
respect of shape and size. A custom-fit bicycle helmet, designed other hand, what seems most important is the experience pro-
based on the individual's head shape, is expected to improve fit and vided, which was defined as the hedonic value (pleasure, fun,
comfort of the bicycle helmet. Even though the most important inspiration and excitement felt during the MC experience), and the
function in wearing a bicycle helmet is to provide head protection creative fulfilment value (accomplishment related to the creative
to the cyclist in the event of an accident, numerous surveys from task of co-designing).
the literature have indicated wearing a helmet is not comfortable The need for better-fitted helmets has been highlighted in the
and the current sizing of helmet does not provide a good fit for the literature since the 1990s (Robinette and Whitestone, 1994).
users (Kotha, 1995; Fogliatto et al., 2012; Westbrook and However, little improvements have been reported for helmets in
Williamson, 1993; Lampel and Mintzberg, 1996). The leading terms of fit (Thai et al., 2015a, 2015b; Ellena et al., 2016). In our MC
reason for this helmet discomfort is that the human head shape and approach, a generic bicycle helmet model is automatically rede-
dimensions are different according to ethnic background, age and signed to fit the customer's head shape. Consequently, the new
gender (Pine, 1993; Da Silveira et al., 2001; Hart, 1995). It was also liner will improve the user's perceived comfort and will add
reported that Asian users experience poor fit because most helmets product value in terms of uniqueness and hedonic values.
currently available on the market are designed according to the size Furthermore, as pointed out by Fiore et al. (2003), the use of 3D
of Western heads (Da Silveira et al., 2001; Gilmore and Pine, 1997). body scanning can contribute to added benefits and may increase
Therefore, based on these helmet fit problems reported in the the customers' willingness to take part in mass customised
literature, we could conclude that the users are not satisfied with products.
the current general size of the bicycle helmet and it does not pro- Market conditions must be appropriate. MC can offer significant
vide the best selection of sizing to improve the helmet fit, comfort competitive advantages over competitors, especially when the
and ergonomic. In addition, the discomfort of a product, in this case, system is original (Kotha, 1995). From a producer's perspective,
due to the helmet fit, is also one of the customer's sacrifices gap value is added by premium prices for mass customised products,
(Hart and C.W., 1995). The bigger the gap would also drive the need increased customer loyalty and improved brand reputation (Piller
of customisation of a product (Gilmore and Pine, 1997; Hart and et al., 2004). All of which could bring significant market shares to
C.W., 1995; Rogers and Peppers, 1993; Whiteley and Hessan, manufacturers, who are willing to embark on the MC production
1996). Other examples of the sacrifices gaps are such as hassle, method.
inconvenience, long queues, product deficiencies, high cost and While individualisation in the garment industry is now recog-
ordering difficulties (Hart and C.W., 1995). A potential solution to nised has a valuable alternative to standardisation (Otieno et al.,
overcome the helmet fit problem for each individual is the mass- 2007; Song and Ashdown, 2012; Lim, 2009), very scarce works on
customisation helmet design approach, where the helmet liner is helmet customisation has been either reported in the literature or
designed according to the size and shape of the head of the indi- initiated by industries. Lui et al (Liu et al., 2008). first attempted to
vidual. This approach could improve the helmet fit and comfort, design custom-fit helmets using a semi-parametric surface
and at the same time satisfy all head shape, regardless of age, modelling tool and 1D anthropometric data (e.g. head circumfer-
gender and ethnic background. However, there have been only ence, head breath, and head length). Their helmets were fabricated
limited studies addressing the mass customisation framework of from only a hard shell (no foam liner), with a simple rounded egg-
bicycle helmet. Most previous studies in the literature focused on shape that offers limited design features. These helmets can be
the materials used for the liner, such as Functionally Graded Foam designed as simple parametric models with just a handful of pa-
Liner (FGF) (Thai et al., 2015a), the dual-material combination of rameters. Conversely, contemporary bicycle helmets have complex
two polyurethane foams (Thai et al., 2015b), and Aluminium hon- free-form shapes with ventilation holes that require advanced
eycomb (ALH) (Ellena et al., 2016), and their influence on the design models. Pandremenos and Chryssolouris (2009) created a
impact performance of the helmet. Other studies have also custom-fit motorcycle helmet liner (again a simple rounded egg-
described new helmet liner designs, such as the liner design using shape design) using a modular design approach and rapid
deformable cones made from Acrylonitrile Butadiene Styrene manufacturing technologies. Although the method proposed could
(ABS), which was proposed by Blanco et al. (Fiore et al., 2003). be applicable to many different customised products other than
In this work, we built our MC framework around the transparent helmets, 3D printing the liner using polyurethane will greatly alter
principle (Gilmore and Pine, 1997), where products are almost fully the shock absorption properties of the helmet. This issue was not
altered to match the needs of each individual (i.e. need for well addressed by the researchers. In 2013, Bell Sports® (Rantoul, Illi-
fitted helmets). However, we keep a modular approach in our nois, USA) launched a Custom-Fit program for two of their motor-
design process where only the inside foam liner of a standard cycle helmet models. Based on a 3D scan of the user's head, Bell
helmet model is altered to fit the customer's head shape and size. In Sports® claimed that the Expanded Polystyrene (EPS) liner is
the following section, we justify the need for custom-fit helmet individually redesigned to fill the void between the head shape and
models by addressing the two market-related factors identified in the shell.
the literature for successful implementation of MC systems (Da Safety standards and certification may be one of the main rea-
Silveira et al., 2001; Fogliatto et al., 2012). sons for the lack of MC systems of helmets. According to interna-
Customer demand for customisation must exist. Customers must tional and national standards, helmets are to be tested on a range of
appreciate the added value of MC products to initiate demands. standard mannequin heads called headforms. They aim to
124 T. Ellena et al. / International Journal of Industrial Ergonomics 64 (2018) 122e133

represent the full range of head dimensions, geometries and shapes generated digital custom-fit helmets using the parametric CAD
within a specific population. Physical helmets of the intended (Computer Aided Design) modeling characteristics of a new generic
design are tested in a set of experiments specified in standards. For bicycle helmet model (Section 2.4). (5) Finally, we evaluated the
instance, Standards Australia uses the tests described in AS/NZS design of the custom-fit helmet models based on the best and
2512:2009 (Australian/New Zealand St, 2009), which follows the worst helmet scenarios using a finite element analysis (FEA)
guidelines and requirements set in AS/NZS 2063:2008 (Australian/ method (Section 2.5).
New Zealand St, 2008). While each customised model of Bell
Sports MC system meets both the DOT standard in the US 2.1. 3D anthropometric survey and clustering algorithm
(Department of Transportation) and the Snell M2015 standard
(Snell Memorial Foundation, 2015), little information on how they The 3D anthropometric database used in this study was pre-
achieve the aforementioned is disclosed (although, the Snell Me- sented in (Perret-Ellena et al., 2015). The survey took place at
morial Foundation® (North Highlands, California, USA) seems to multiple sites around the metropolitan area of Melbourne, Australia
have new criteria for the certification of customised motorcycle and spanned over a 9-month period throughout 2014 (April to
helmets (Snell Memorial Foundation, 2013)). Certainly, certifying December). A total of 222 cyclists, age 18 years or above, had their
every customised design using multiple physical helmets (e.g. 10 head digitized using a portable white light 3D scanner. The hair
specimens are required in (Australian/New Zealand St, 2008)) thickness under the wig cap compression was measured and
would not be cost and time effective. We introduce an innovative evaluated during the post-processing stage to match the exact head
approach for the automatic ‘certification’ of custom-fit bicycle shape of the participants.
helmet models. The automatic certification is to confirm the fit We introduced (Ellena et al., 2018) an innovative data mining
accuracy and impact energy absorption level of the customised method for the classification of head shapes. Through a modified
helmets. hierarchical clustering algorithm, distance metrics between pairs of
Although MC is defined as the use of both computer-aided registered head scans (i.e. point set registration method for com-
design and manufacturing systems to produce a custom output, parison of 3D shapes (Allen et al., 2003)) were calculated and
this study focused exclusively on the design aspect of our MC implemented in a step-by-step process, where clusters were
framework of bicycle helmets. More specifically, we tried to answer created one after another (instead of simultaneously in a conven-
the following central research question: ‘How one can automati- tional hierarchical algorithm) in an enhanced manner using simi-
cally create commercially ready, custom-fit digital 3D models of larity criteria measures. When applied to the 3D anthropometric
bicycle helmets based on 3D anthropometric data?’ Manufacturing database of Australian cyclists, four groups of head shapes, denoted
guidelines for future work are however provided in the conclusion as cluster N
1 to N
4, were computed.
section. We develop and demonstrate our approach around a case
study where customised helmet models are created for Australian 2.2. Minimum, mean and maximum head shape representations
cyclists. We evaluate the accuracy of fit of the custom-designed
models using objective assessments and expected the custom-fit Stemming from these computed clusters, four new headform
helmet models to improve helmet fit and comfort of the current models were generated using the mean shape of all participants
helmet designs on the market. within the same group (Ellena et al., 2017). As we aimed to create
custom-fit helmet models that comply with the relevant safety
2. Materials and methods standards, we made sure that only small variations of the liner
thickness were permitted during the customisation process. This
We built our design and certification method through a five-step was to ensure that the shock absorptions characteristics of the
procedure. (1) We first partitioned the population of interest (i.e. helmets were not dramatically altered.
Australian cyclist) based on head shape and size (Section 2.1). More Considering the above, we created the Maximum Head Shape
specifically, individuals with similar head shapes were grouped surface (MaH) that can be accommodated in a group. During the
together to create a new sizing system of the head. (2) We then customisation process, this surface can never be violated on the
generated novel headform models to represent the minimum, external side. This is the worst case scenario (i.e. the smallest liner
mean and maximum head shapes of each group (Section 2.2). These thickness at all locations around the helmet) of a specific helmet
shapes were important as they represented the best and worst model size. Similarly, we defined the Minimum Head Shape surface
helmet scenarios (thickness of the foam) within each cluster. (3) (MiH). This is the best case scenario (i.e. the largest liner thickness
Subsequently, we developed a 3D head shape classifier where new at all locations around the helmet). This surface can never be
customers were assigned to one of the pre-computed groups for violated on the internal side as we seek to limit the maximum
customisation (Section 2.3). This approach minimized the varia- weight for each helmet size.
tions of the liner thickness within each group size. (4) Next, we The two shapes were constructed using Boolean operations

Fig. 1. Union Boolean operation between the first two participants in cluster N
3. The maximum shape is kept. The MaH is created by combining every individual in a cluster using
this operation.
 T. Ellena et al. / International Journal of Industrial Ergonomics 64 (2018) 122e133 125

Fig. 2. Intersect Boolean operation between the first two participants in cluster N
3. The minimum shape is kept. The MiH is created by combining every individual in a cluster
using this operation.

Fig. 3. Final MaH (top row) and MiH (Bottom row) for our clusters N
1 to N
4 (from left to right).

within Geomagic Studio 12® (rock Hill, SC, USA) and the point set  Point Set Registration: The new participant's head shape is
registration process described in (Ellena et al., 2018). The Union and registered according to the procedures highlighted in (Ellena
Intersect tools were used to compute the MaH and MiH surfaces, et al., 2018).
respectively, as shown in Figs. 1 and 2 for two participants in cluster  Initialization: The Head Covering Points (HCPs) (i.e. Points of the
N
3. Fig. 3 shows the final MaH and MiH surfaces for the four participant's head scan that should be under helmet protection)
computed groups. The faces features (e.g. nose, eyes, and ears) were positions are recorded for the new participant and the min and
kept for visualisation only. Only the shape of the head that should max head shapes (MaH and MiH) of the four computed clusters.
be under helmet protection was analysed and processed during the  Space Partitioning: K-d trees are constructed for the thousands
clustering process (Ellena et al., 2016) (green area in Fig. 8). of vertices defining the MaH and MiH shapes.
 Nearest Neighbor Search: A point search algorithm is run to find
the closest correspondences between the HCPs of the partici-
pant's head and the MaH and MiH shapes.
 Classification: Distance metrics are computed for these corre-
2.3. 3D head shape classifier spondences to assess whether or not the tested head shape is
inside the min and max limits of each cluster's space.
In machine learning, clustering techniques are considered as  Optimized Cluster Selection: If a participant belongs to more
unsupervised procedures where objects are grouped into cate- than one group, a set of parameter measures assess which of the
gories based on some measure of intrinsic similarity. Conversely, clusters' candidate will produce the thinnest (and hence light-
supervised learning is the process of classifying new observations est) customised helmet for the head shape considered.
into one of the available categories. Such algorithms that categorize
objects are known as classifiers in concrete implementations.
In our project, we developed a novel 3D Head Shape Classifier 2.3.1. Nearest Neighbor Search for Head Covering Points
that categorizes new customers into one of the computed clusters. correspondences
Participants could also belong to none of these four clusters. Euclidean distance metrics were used to assess if the tested
The 3D Head Shape Classifier works as follows: head shapes were within the boundary limits of one of the four
126 T. Ellena et al. / International Journal of Industrial Ergonomics 64 (2018) 122e133

Fig. 4. Example of distance metrics (di ) between two registered meshes at seven Head Covering Points. The smallest distances are not necessary between the same labelled points
(e.g.; d3 0 and d6 0 ).

computed clusters. However, instead of computing these distances (2) The distances between these correspondences (MaH and
between the labelled HCP, we decided to first update the pre- MiH) at each HCP of the tested head shape were then
liminary vertices correspondences using K-d trees and Closest recorded (green lines in Figs. 6 and 7).
Point Search Algorithm. Even though the point set registration (3) For each HCP of the tested shape, a point was considered
process ensures that similar deformations apply within triangles outside the boundary limit (i.e., Mah and MiH) of a cluster if
located in the same region of the head (stiffness term of the cost any of the two calculated distances in (1) (i.e. grey lines) was
function in (Ellena et al., 2018)), it is not guaranteed that the larger than the distance computed in (2) (i.e. green lines).
smallest distance from one vertex i of a registered mesh to another (4) If more than 99% of the HCPs were located inside the
will be located at this specific point i (Fig. 4). boundary limits, the shape was considered to belong to the
Preliminary correspondences between every vertex of the tested tested cluster.
head mesh and the vertices of both the MaH and MiH are found (5) If a head shape belonged to more than one cluster, the se-
using Nearest Neighbor Search (NNS) (Yianilos, 1993) (e.g.; Fig. 5). lection was made based on the sum of distances between the
Formally, the NNS problem is stated as: given a set of points HCPs of the head and the MaH. The cluster with the smallest
u2U (i.e.; the MaH or MiH meshes), and a set of query points v2V sum of distances was selected.
(i.e.; the tested head mesh), find for all v the closest point to U.
Perceptibly, the simplest solution is to compute the distance for 2.4. Parametric CAD modeling
each query point to all the points in the MaH and MiH meshes and
keep track of the closest point. However, for each NNS, this linear The next step in our MC system was to use the scanned head
search has a running time of Oð3nÞ where n is the number of shape as an input element for the design modification of a generic
vertices. bicycle helmet model. However, we had to first construct a surface
To speed up the NNS problem, we used a space-partitioning model of the head since it was still defined by a polygon mesh (i.e. a
method called K-d tree (K-dimensional tree) (Bentley, 1975). The point cloud from the scanned data). The surface had to be accurate
algorithm has a Oðlog nÞ average complexity in the case of enough to capture the head shape characteristics of the individual,
randomly distributed points (Curry and Schoenberg, 1966). but in the meantime had to be smooth enough to avoid small
wrinkle on the final design model. That is, the created surface had
2.3.2. Classification procedure to curve when necessary, but not bend in all places to try to copy
The classification was based on the following five steps: perfectly every point position of the mesh (noise might be present
in the data and should be avoided).
(1) The Euclidean distance metrics between the HCPs of the
tested head shape and the vertex correspondences in the 2.4.1. Curve and surface fitting
MaH and MiH shapes were first computed (grey lines in The head surface, named HPP, was built following the definition
Figs. 6 and 7). of the Head Protection Proportion area in (Ellena et al., 2016). We

Fig. 5. Point correspondences (grey lines) between the tested head shape and the MaH and MiH of one cluster.
 T. Ellena et al. / International Journal of Industrial Ergonomics 64 (2018) 122e133 127

Fig. 6. An example of a tested head shape located inside the boundary limit of one of the computed cluster. Grey and green lines are Euclidean distance metrics. (For interpretation
of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Fig. 7. An example of a tested head shape that is considered outside the boundary limit of one of the computed cluster. Grey and green lines are Euclidean distance metrics. (For
interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

built the HPP using two geometric elements; the Head Covering the model that should be common for all customised helmets
Curve (HCC) and the Head Covering Surface (HCS). See Fig. 8 for a within the group. In our example, the outside was created using
graphic representation of the three geometric elements on a two free-forms surfaces with G2 curvature continuity (Fig. 9). The
random mannequin head. B-Spline curves and B-Spline surfaces surfaces were primary defined around the MaH shape (red dash in
(Piegl and Tiller, 2012) were used for the HCC and HCS, respectively. Fig. 9) with an offset distance spanning between 22 and 40 mm. The
Many different methods exist to fit B-Spline curves and surfaces bottom boundary limit (green in Fig. 9) was created using the HCC
on polygon meshes. For the HCC, we approximated 50 data points of MaH (red in Fig. 9), MiH (blue in Fig. 9), and the 108 individuals
using B-Spline curve fitting (Piegl and Tiller, 2012) and the closing classified in cluster N
1 in (Ellena et al., 2018) (white in Fig. 9).
point derivatives based on Peigle's algorithm (Piegl and Tiller, Next, we created the inside surface of the generic helmet using
2000). For the HCS, we approximated the 13000 Head Covering the MiH surface as an input (Fig. 10). This is the maximum liner
Points provided by the registration method and the HPP area using thickness a custom-fit helmet can inherit after the customisation
an iterative Genetic Algorithm (GA) approach introduced by Ga lvez process.
lvez et al., 2012). The HPP surfaces were also generated
et al. in (Ga We finally designed the vents of the generic helmet model using
for the MaH and MiH elements of our four computed clusters. a five-row system of large openings at the top (Fig. 11). Four aper-
tures were also added at the back using more rounded and elon-
2.4.2. Design gated shapes. The reinforcement features (yellow surfaces in
We demonstrated our helmet custom-fit design method on Fig. 12) were positioned in a certain way to avoid interferences with
group N
1. The process involved designing a generic helmet model a customer's head shape during the customisation process. This
which was then modified internally (i.e. the shape and thickness of was achieved by keeping the geometric elements of all the rein-
the foam liner) to fit the individual's head shapes. The design forcement features above the MaH surface (red in Fig. 12).
process was split into two phases; the standardisation design, and The second design phase in our custom-fit design framework
the customisation design. was the customisation process, where the head shape of an indi-
In the standardisation phase, we designed the rigid surfaces of vidual (i.e. HPP surface) was used as an input for the modification of
128 T. Ellena et al. / International Journal of Industrial Ergonomics 64 (2018) 122e133

Fig. 11. The generic bicycle helmet model for group N
1.

added to finalise the model.

The helmet shell was also defined as a standard component
(Fig. 14). It was designed using most of the outer surfaces of the
helmet liner. The bottom boundary of the side surface was defined
Fig. 8. Yellow ¼ HCS, blue ¼ HCC, brown ¼ template mesh, green ¼ HPP. On the left, using the maximum HCC in the group, as it represented the highest
the computed HCS (defined as a B-Spline surface) extends below the HCC (defined as a helmet boundary position for any individual in the group.
B-Spline curve) and cover the measured data points from the head area that should be The helmet model was designed using CATIA V6 CAD software
under helmet protection. The right image shows the final HPP surface that serves as an mes®, Ve
lizy-Villacoublay, France) and the Knowl-
(Dassault Syste
input element during the customisation design process. (For interpretation of the
references to colour in this figure legend, the reader is referred to the Web version of
edgeware workbenches used in the parametric system. This
this article.) approach permitted the development of a fully parametric, inte-
grated system where the helmets were automatically customised
based on the customer ID.

2.5. Finite element analysis for helmet certification

We developed a novel certification method for the custom-fit

helmet models. The procedure rests upon the idea that every sin-
gle customised helmet created within a group should be safe to use
if the worst- and best-case helmets (i.e. helmets based on the MaH
and MiH surfaces) comply with the relevant safety standards.
Consequently, the physical tests required for certification should
Fig. 9. Initial outside surface of the generic customised helmet model for cluster N
1. only be performed for the two extreme models of each group. We
The side view on the right shows the outline profiles of the MaH (red dash) and the demonstrated this assumption by using the validated drop impact
MiH (dash blue) surfaces, and the HCC of the MaH (red), the MiH (blue) and the 108 test simulation models that complied with the Australian standards
individuals in the cluster (white). The green line is the bottom boundary limit of the (Australian/New Zealand Standard, 2007). The simulation was
generic model. (For interpretation of the references to colour in this figure legend, the
performed in Abaqus 6.14 (Dassault Syste mes®, Velizy-Villacoublay,
reader is referred to the Web version of this article.)
France).
Geometric data of the helmet liners, shell, headforms and flat
anvil were imported to the software. The helmet shell was meshed
using triangular S3R linear shell elements, and the liners were
meshed with C3D10M modified quadratic tetrahedral elements.
Distortion control was also applied to all elements of the liners to
avoid excessive distortion. Meanwhile, C3D10M tetrahedral ele-
ments were used to mesh the Australian standard J-headform
(Australian/New Zealand Standard, 2009), and C3D8R, an 8-node
linear brick element, was chosen for the flat anvil. The headform
selection was based on deviation analyses between the computed
mean shape of cluster N
1 and the standard Australian headforms.
The most resembling headform, i.e. J-headform, was then selected
Fig. 10. The inside design of the generic helmet based on the MiH surface (blue). Right for the simulation (see (Ellena et al., 2017) for full details and an-
is a section view along the mid-plane. (For interpretation of the references to colour in alyses) as part of the requirements set by Australian standards
this figure legend, the reader is referred to the Web version of this article.)
(Australian/New Zealand Standard, 2007).
The helmet liners were modeled using an isotropic crushable
foam material with a density of 65 kg/m3. Volumetric hardening
the generic model presented above. The procedure consisted of parameters such as the ratios of the initial yield pressures in hy-
combining the HPP and HCC shapes to create a new inner surface of drostatic tension and compression and the uniaxial compressive
the helmet liner. A simple split operation was implemented. Fig. 13 data of expanded polystyrene (EPS) were taken from literature
shows an example of the process where the head shape of an in- (Mills and Gilchrist, 2008). Other material properties of helmet
dividual (orange) was used to generate the customised helmet. components such as shell and J-headform were extensively
Dress-ups features such as fillets, chamfers and drafts were then described in [Mustafa El Bakri et al, 2015; Mustafa et al, 2015].
 T. Ellena et al. / International Journal of Industrial Ergonomics 64 (2018) 122e133 129

Fig. 14. Shell of the generic helmet model (blue). (For interpretation of the references
to colour in this figure legend, the reader is referred to the Web version of this article.)

The customized helmet was tested on three impact locations:

crown, front and sides (as shown in Fig. 15), replicating the similar
positions to those in the experimental impact tests. The impact
velocity of the helmet and headform were set at 5.44 msˉ1, which
was obtained from experimental drop impact tests [Mustafa et al,
Fig. 12. The design of the reinforcement features of the generic bicycle helmet model. 2015].
P1, P2, and P3 are three planes passing through the main aerations of the helmet. The
three section views along these planes are represented. The green surfaces are the liner
sections intersecting the associated plane. These surfaces can be trimmed down during 3. Results and discussion
the customisation process. The yellow surfaces are the opening reinforcements. They
are fixed and cannot be changed. The pink contours are the reinforcement sections at 3.1. Classification results
the specific cutting plane. Similarly, the red dash lines represent the intersection of the
planes with the MaH surface. As shown in the graphics, minimum distance values were
kept between the red and pink elements to allow a slight gap between the top features Fifteen participants from the 3D Anthropometric database of
of the customised helmet and the head of any possible individual in the group. (For Australian cyclists (Perret-Ellena et al., 2015) were chosen for the
interpretation of the references to colour in this figure legend, the reader is referred to testing of the 3D Head shape Classifier. The selected participants
the Web version of this article.)
were not part of the clustering algorithm study (Ellena et al., 2018).
Table 1 presents the results. Fourteen participants were classified in
one of the four clusters (93.3%), including four in multiple clusters.
Fig. 16 shows the cross-sectional graphic representation of a par-
ticipant's head mesh in between the two limits of cluster N
3.
A large percentage of the tested individuals (53%) were assigned
to group N
1, which is the most common head shape in our
Australian sample. This is consistent with our clustering results in
(Ellena et al., 2018) where 54% of the sample were categorized in
this group.

3.2. Customised design

Using the MC procedures highlighted in section 2.4.2, custom-fit

helmet models were created for the 116 individuals classified in
group N
1 (108 from the clustering algorithm in (Ellena et al.,
2018), and 8 form the 3D Head Shape Classifier in Table 1). Fig. 17
shows an example of custom-fit helmet models created for five of
these individuals. A cross-sectional view of the five helmet liners is
Fig. 13. Customisation process using the HPP and HCC shapes of an individual scanned presented in Fig. 18 where the different thicknesses resulting from
3D anthropometric head data (orange). (For interpretation of the references to colour the customisation process can be evaluated.
in this figure legend, the reader is referred to the Web version of this article.) Table 2 presents the helmet liners statistics in terms of volumes
and thicknesses for the 116 custom-fit designs tested. The data are
compared to the best-case (maximum thickness) and worst-case
Penalty contact property with friction coefficient of 0.4 was adop- (minimum thickness) helmets generated from the MaH and MiH
ted for interactions between all surface contacts in the simulation. surfaces. Results demonstrate that the custom-fit models lie in
Meanwhile, tie constraint was applied between the inner surfaces between these limits with values spreading inside the available
of the shell and the outer surfaces of the liner to simulate in-mold ranges. Although the ratio of individuals classified in group N
1 is
bonding, which is common for bicycle helmet (Mills and Gilchrist, large, the customised helmets generated seem to be fairly similar in
2008). Homogenous shell section with corresponding thickness of terms of volume. For example, the mean liner volume is 1798 cm3,
0.40e0.45 mm was assigned to helmet shell. The anvil and head- while the volume of the minimum and maximum were 2116 cm3
form were defined as rigid bodies, and the bottom face of the anvil and 1484 cm3 respectively. Using a standard foam density of 65 kg/
was fixed in every degree of freedom. m3, we show that the difference in the liners’ weight, between the
130 T. Ellena et al. / International Journal of Industrial Ergonomics 64 (2018) 122e133

Fig. 15. Impact locations of the customized helmet: side, front and top.

Table 1
Group classification for the 15 participants tested. A “X” means that the head shape was classified in the cluster. The red “X” is the final selection.

Participants Cluster № 1 Cluster № 2 Cluster № 3 Cluster № 4

№1 X
№2 X
№3 X
№4 X X
№5 X X
№6 X
№7
№8 X
№9 X X
№ 10 X X
№ 11 X
№ 12 X X
№ 13 X
№ 14 X
№ 15 X
TOTAL 8 2 2 2

clustering algorithm (Ellena et al., 2018). Yet, a high increase in

production cost is to be expected.

3.3. Design validation

A validated simulation model was used to conduct virtual drop

impact tests for five custom-fit helmet models created in group N

1 (Fig. 17). In addition, the worst-case and best-case scenarios were
also tested.
According to the AS/NZS standard for bicycle helmet, a bicycle
helmet should obtain peak linear acceleration below 250 g to be
considered as a safe helmet (Australian/New Zealand Standard,
2007). In the drop test simulation, the peak linear accelerations
(PLA) of the seven helmets were plotted against time at three main
impact locations, i.e. top, front and side (Fig. 19). Is it clear from the
graphs that all custom-fit helmets tested recorded PLAs below the
safety limit value of 250 g. For the three impact locations, the
worst-case helmets recorded the highest PLAs, while the lowest
PLAs were obtained by the best-case helmets.
Fig. 16. Cross-Section view of a participant (bleu) sandwiched in between MaH and
This is consistent with the known fact that PLA is highly
MiH shapes (orange). (For interpretation of the references to colour in this figure
legend, the reader is referred to the Web version of this article.) dependent on the helmet thickness. As shown in Table 2, the worst-
case helmet has the thinnest liner and the best-case helmet has the
thickest liner. The PLAs for the five helmet models were in between
two extreme cases, is only 23 g. However, we notice that the liner the best-case and the worst-case helmets at the three locations.
thicknesses could be fairly dissimilar between the created models Again, this result is also consistent because the thicknesses of these
(up to 14 mm difference). This could create annoyance to some custom-fit helmet models were in between the worst-case and the
customers that would end up with bulky helmets on their head. A best-case helmets.
simple solution would be to implement more standard sizes to the As a result, we conclude that all the customised helmets created
MC framework by modifying the stopping parameter of our through our MC framework comply with the relevant safety
 T. Ellena et al. / International Journal of Industrial Ergonomics 64 (2018) 122e133 131

Fig. 17. Examples of customised helmet designs for five individuals belonging to group N
1.

The suitability of the index to assess the helmet fit was validated

using cyclist's subjective assessments of three commercially avail-
able helmets (Ellena et al., 2016).
Most of the participants included in the HFI study (Ellena et al.,
2016) were also involved in the grouping analysis in (Ellena et al.,
2018) and the custom-fit design study of bicycle helmets pre-
sented in section 3.2. This configuration allowed us to validate the
high fit accuracy of our customised helmets compare to the three
Fig. 18. Cross-sectional view of the five customised helmet models from Fig. 17. commercially available models studied.
We found 61 individuals who took part to the HFI study and also
belong to group N
1. The first row of Table 3 presents the HFI
statistics of the 61 custom-fit helmet models generated for these
standards when the liner thicknesses are between the worst- and
specific individuals. The raw head mesh data generated by the 3D
best-case helmets. This statement is true when the worst case
scanner were used for the analyses (Fig. 20) to produce results as
scenario pass the impact tests requirements described in
accurate as possible (i.e. using the registered head meshes or the
(Australian/New Zealand Standard, 2007). Although the best-case
HPP surfaces would have produced much higher HFIs values). The
model is not essential to validate our method, we added it to the
next three rows of Table 3 show the same HFI statistics for the three
simulation to ensure that all the custom-fit models generated for a
commercially available helmets.
group hold similar impact characteristics. For example, we know
As a result, using three paired-samples t-test, we were able to
that a drop test at the front location in one of our customised
determine that the HFI values of our customised helmets were
helmet in group N
1 would have a PLA between 108.4 g and
statistically significant higher than the three commercially avail-
174.3 g.
able models (p < 0:0005). The interpretation of the Cohen's d values
showed a strong practical difference (d > 0:8 for the three tests).
3.4. Evaluation of helmet fit
Since the HFI has been shown to predict subjective feeling
regarding fit (Ellena et al., 2016), we conclude, as expected, that
The accuracy of fit for the customised helmets created was
custom-fit design of helmets improve the accuracy of fit by a sig-
assessed using an objective evaluation method called the Helmet
nificant margin.
Fit Index (HFI) (Ellena et al., 2016). Based on 3D anthropometry,
reverse engineering techniques and computational analysis
methods, the index provides a fit score on a scale ranging from
4. Conclusion and future work
0 (excessively poor fit) to 100 (perfect fit).
The three essential parameters used in the HFI formula are the
A novel approach for the design of bicycle helmets was pre-
Standoff Distance SOD (average distance between the head and the
sented in this work. Based on 3D anthropometry, we automatically
helmet's liner), the Gap Uniformity GU (measure of SOD disper-
generated custom-fit digital helmet models well-fitted to the head
sion), and the Head Protection Proportion HPP (measure of the
shape of the wearer. This was achieved through the introduction of
head surface area percentage under helmet protection).
multiple innovative procedures.
8   Using a clustering algorithm, we first classified a subset of the
>
> jSOD  6j 0:12GU
>
< 100*exp 0:13   for 4 > SOD> 8 studied population according to their similar head shapes. This
15 HPP
HFI ¼   categorization enabled the creation of four new standard helmet
>
> 0:12GU
>
: 100*exp  for 4  SOD  8 sizes for the implementation of our customisation platform. The
HPP idea was to perform the customisation at the group level and follow
a modular approach by only personalising the inside surfaces of the

Table 2
Statistics of customised helmet liners in. Sample size ¼ 116.

Statistic Mean ± SD Minimum Maximum Best Case Worst Case

Helmet liner volume (cm3 Þ 1798 ± 52 1640 1988 2116 1484

Mean liner thickness (mm) 36.6 ± 1.8 31.4 39.7 44.3 29.4
Minimum liner thickness (mm) 30.3 ± 2.3 25.1 38.8 38.8 25.1
Maximum liner thickness (mm) 53.2 ± 3.6 46.7 61 61 46.7
132 T. Ellena et al. / International Journal of Industrial Ergonomics 64 (2018) 122e133

Fig. 19. Peak linear acceleration of five custom-fit helmets (grey), the best-case helmet (green), and the worst-case helmet (red). The simulations were performed at three locations,
namely the front (top graph), the top (middle graph), and the side (bottom graph). (For interpretation of the references to colour in this figure legend, the reader is referred to the
Web version of this article.)

Table 3
Custom-fit helmets assessment study - Summary Statistics e Data are mean (95% CI) e Sample size is 61.

SOD (mm) GU (mm) HPP HFI

Customised Helmet group N
1 2.49 (2.29, 2.68) 0.93 (0.84, 1.00) 0.96 (0.96, 0.96) 80.1 (79.3, 80.8)
Helmet A 9.0 (8.8, 9.4) 4.0 (3.8, 4.1) 0.83 (0.82, 0.84) 52.8 (50.9, 54.6)
Helmet B 8.3 (7.9, 8.7) 3.7 (3.5, 3.9) 0.85 (0.84, 0.86) 55.7 (53.6, 57.7)
Helmet C 9.9 (9.6, 10.4) 4.2 (4.1, 4.4) 0.83 (0.82, 0.83) 48.2 (46.0, 50.4)

helmet foam liner. In order to do so, we created a generic helmet helmet model and the transformed data of the digitized head.
model for each group based on the minimum and maximum head We also demonstrated that the customised helmets created
shapes present in a group. with our method complied with the relevant drop impact test
The customisation process involved three steps: (i) digitization standards if their liner thicknesses were within specific boundary
of a customer's head shape, (ii) categorization into one of the four limits. The limits were set by the best and worst case helmets for
predefined group, and (iii) modification of the inside shape of the each group. This finding should significantly help with the potential
generic helmet model based on the customer's head surface. We commercialisation of our custom-fit helmet models where only the
achieved step (ii) using supervised learning techniques, and sub- worst-case helmet models of each group will need to be physically
sequently developed a method named 3D head shape classifier. To tested.
the authors' best knowledge, this is the first time that 3D head Finally, in this study, we proved that the fit accuracy of our
scans are objectively classified into predefined groups of in- models was significantly increased compared to three commer-
dividuals based on their similar head shapes. Step (iii) was cially available helmets. Such an improvement should hold
accomplished through standard 3D modelling techniques, where a competitive advantage for producers who are willing to embark on
simple split operation was performed between the created generic the MC production of bicycle helmets.
 T. Ellena et al. / International Journal of Industrial Ergonomics 64 (2018) 122e133 133

Ellena, Thierry, Subic, Aleksandar, Mustafa, Helmy, Pang, Toh Yen, 2018. A novel
hierarchical clustering algorithm for the analysis of 3D anthropometric data of
the human head. Comput. Aided Des. Appl. 15 (1). https://doi.org/10.1080/
16864360.2017.1353727.
Ellena, Thierry, Skals, Sebastian, Subic, Aleksandar, Mustafa, Helmy, Pang, Toh Yen,
2017. 3D digital headform models of Australian cyclists. Appl. Ergon. 59 (Part A),
11e18. ISSN 0003-6870. https://doi.org/10.1016/j.apergo.2016.08.031.
Fiore, A.M., Seung-Eun, L., Kunz, G., 2003. Psychographic variables affecting will-
ingness to use body-scanning. J. Bus. Manag. 9 (3), 271.
Fogliatto, F.S., da Silveira, G.J., Borenstein, D., 2012. The mass customization decade:
an updated review of the literature. Int. J. Prod. Econ. 138 (1), 14e25.
Galvez, A., Iglesias, A., Puig-Pey, J., 2012. Iterative two-step genetic-algorithm-based
method for efficient polynomial B-spline surface reconstruction. Inf. Sci. 182 (1),
56e76.
Gilmore, J.H., Pine, B.J., 1997. The four faces of mass customization. Harv. Bus. Rev.
75, 91e101.
Hart, C.W.L., 1995. Mass customization: conceptual underpinnings, opportunities
and limits. International Journal of Service Industry Management, 6 (2), 36e45.
Hart, C.W., 1995. Mass customization: conceptual underpinnings, opportunities and
limits. Int. J. Serv. Ind. Manag. 6 (2), 36e45.
Kotha, S., 1995. Mass customization: implementing the emerging paradigm for
competitive advantage. Strat. Manag. J. 16 (S1), 21e42.
Lampel, J., Mintzberg, H., 1996. Customizing customization. MIT Sloan Manag. Rev.
38 (1), 21.
Lim, H., 2009. Three Dimensional Virtual Try-on Technologies in the Achievement
and Testing of Fit for Mass Customization.
Liu, H., Li, Z., Zheng, L., 2008. Rapid preliminary helmet shell design based on three-
dimensional anthropometric head data. J. Eng. Des. 19 (1), 45e54.
Merle, A., Chandon, J.-L., Roux, E., 2008. Understanding the perceived value of mass
customization: the distinction between product value and experiential value of
co-design. Rech. Appl. Market. 23 (3), 27e50.
Fig. 20. Deviation analysis of a participant's customised helmet. SOD is 1.99 mm, GU is Merle, A., et al., 2010. Perceived value of the Mass-Customized product and mass
0.87 mm, HPP is 0.95 (HFI ¼ 78.1). customization experience for individual consumers. Prod. Oper. Manag. 19 (5),
503e514.
Mills, N., Gilchrist, A., 2008. Finite-element analysis of bicycle helmet oblique im-
pacts. Int. J. Impact Eng. 35 (9), 1087e1101.
Even though the implementation of advanced manufacturing Mustafa El Bakri, H., et al., 2015. Finite element bicycle helmet models develop-
technologies is essential for successful development of MC systems, ment. In: Destech 2015. Elsevier BV.
it was not described in this study. Promising techniques such as Mustafa, H., et al., 2015. Finite element bicycle helmet models development. Pro-
cedia Technology 20, 91e97.
additive manufacturing (AM) (e.g. 3D printing) could be considered
Otieno, R., Apeagyei, P.R., Otieno, R., 2007. Usability of pattern customising tech-
for this implementation. However, 3D printing the helmets directly nology in the achievement and testing of fit for mass customisation. J. Fash.
is not an option at the moment. This is because the mechanical Mark. Manag.: Int. J. 11 (3), 349e365.
properties of the available materials differ significantly from the Pandremenos, J., Chryssolouris, G., 2009. Modular product design and custom-
ization. In: Proceedings of the 19th CIRP Design Conferenceecompetitive
well-known foam material that have been used extensively. One Design. Cranfield University Press.
option that the authors intend to present in the future, is the Perret-Ellena, T., et al., 2015. 3D anthropometric investigation of head and face
combination of AM with common molding techniques. In this characteristics of Australian cyclists. In: 7th Asia-Pacific Congress on Sports
Technology: Impact of Technology on Sport, APCST. Barcelona, Spain.
proposal, interchangeable inserts would be incorporated inside a Piegl, L.A., Tiller, W., 2000. Least-squares b-spline curve approximation with arbi-
generic helmet mold. The inserts would be designed based on the tary end derivatives. Eng. Comput. 16 (2), 109e116.
customised helmet models and manufactured using AM. Piegl, L., Tiller, W., 2012. The NURBS Book. Springer Science & Business Media.
Piller, F.T., Moeslein, K., Stotko, C.M., 2004. Does mass customization pay? An
economic approach to evaluate customer integration. Prod. Plann. Contr. 15 (4),
References 435e444.
Pine, B.J., 1993. Mass customizing products and services. Plann. Rev. 21 (4), 6e55.
Addis, M., Holbrook, M.B., 2001. On the conceptual link between mass custom- Robinette, K.M., Whitestone, J.J., 1994. The need for improved anthropometric
isation and experiential consumption: an explosion of subjectivity. J. Consum. methods for the development of helmet systems. Aviat Space Environ. Med. 65
Behav. 1 (1), 50e66. (5 II), A95eA99.
Allen, B., Curless, B., Popovi c, Z., 2003. The space of human body shapes: recon- Rogers, M., Peppers, D., 1993. The One to One Future. Piatkus, USA.
struction and parameterization from range scans. In: ACM SIGGRAPH 2003 Snell Memorial Foundation, 2013. Addendum to Snell Standards for Custom Fitted
Papers. SIGGRAPH '03. Helmets.
Australian/New Zealand Standard, AS/NZS 2063:2008 Bicycle Helmets, 2008. Snell Memorial Foundation, 2015. M2015 Standard for Protective Headgear for Use
Australian/New Zealand Standard, AS/NZS 2512:2009 Methods of Testing Protective with Motorcycles and Other Motorized Vehicles.
Helmets, 2009. Song, H.K., Ashdown, S.P., 2012. Development of automated custom-made pants
Australian/New Zealand Standard, 2007. AS/NZS 2512.3.1:2007 Methods of testing driven by body shape. Cloth. Text. Res. J. 0887302X12462058.
protective helmets. In: Determination of Impact Energy Attenuation - Helmet Spira, J.S., 1993. Mass customization through training at lutron electronics. Plann.
Drop Test. Rev. 21 (4), 23e24.
Australian/New Zealand Standard, 2009. AS/NZS 2512.1:2009 Methods of Testing Squire, B., et al., 2004. Mass customization: the key to customer value? Prod. Plann.
Protective Helmets in Part 1: Definitions and Headforms. Contr. 15 (4), 459e471.
Bentley, J.L., 1975. Multidimensional binary search trees used for associative Thai, K.T., McIntosh, A.S., Pang, T.Y., 2015. Bicycle helmet size, Adjustment, and
searching. Commun. ACM 18 (9), 509e517. stability. Traffic Inj. Prev. 16 (3), 268e275.
Curry, H.B., Schoenberg, I.J., 1966. On Polya frequency functions IV: the fundamental Thai, K.T., McIntosh, A.S., Pang, T.Y., 2015. Factors affecting motorcycle helmet use:
spline functions and their limits. J. Anal. Math. 17 (1), 71e107. size selection, stability, and position. Traffic Inj. Prev. 16 (3), 276e282.
Da Silveira, G., Borenstein, D., Fogliatto, F.S., 2001. Mass customization: literature Westbrook, R., Williamson, P., 1993. Mass customization: Japan's new frontier. Eur.
review and research directions. Int. J. Prod. Econ. 72 (1), 1e13. Manag. J. 11 (1), 38e45.
Davis, S.M., 1989. From “future perfect”: mass customizing. Plann. Rev. 17 (2), 16e21. Whiteley, R., Hessan, D., 1996. Customer-centred growth: five strategies for building
Department of Transportation, FMVSS 218 Federal Motor Vehicle Safety Standard competitive advantage. Manag. Serv. Qual.: Int. J. 6 (5), 47e52.
#218, Motorcycle Helmets. Yianilos, P.N., 1993. Data structures and algorithms for nearest neighbor search in
Ellena, T., et al., 2016. The Helmet Fit Index e an intelligent tool for fit assessment general metric spaces. In: SODA.
and design customisation. Appl. Ergon. 55, 194e207.