The objective of this study was to quantify head-and-face shape variations of U.S. civilian workers using modern methods of shape analysis. The purpose of this study was based on previously highlighted changes in U.S. civilian worker head-and-face shape over the last few decades – touting the need for new and better fitting respirators – as well as the study's usefulness in designing more effective personal protective equipment (PPE) – specifically in the field of respirator design. The raw scan three-dimensional (3D) data for 1169 subjects were parameterized using geometry processing techniques. This process allowed the individual scans to be put in correspondence with each other in such a way that statistical shape analysis could be performed on a dense set of 3D points. This process also cleaned up the original scan data such that the noise was reduced and holes were filled in. The next step, statistical analysis of the variability of the head-and-face shape in the 3D database, was conducted using Principal Component Analysis (PCA) techniques. Through these analyses, it was shown that the space of the head-and-face shape was spanned by a small number of basis vectors. Less than 50 components explained more than 90% of the variability. Furthermore, the main mode of variations could be visualized through animating the shape changes along the PCA axes with computer software in executable form for Windows XP. The results from this study in turn could feed back into respirator design to achieve safer, more efficient product style and sizing. Future study is needed to determine the overall utility of the point cloud-based approach for the quantification of facial morphology variation and its relationship to respirator performance.

The National Institute for Occupational Safety and Health (NIOSH) is continuing research efforts to better understand the relationship between anthropometric factors and respirator face-piece fit. Zhuang et al. have demonstrated that anthropometric data collected from military personnel in the 1960s are no longer reflective of the head-and-face anthropometric distribution of the current U.S. work force (

An additional method for characterizing anthropometric data-sets is 3D point cloud analysis. Three-dimensional point cloud shape variation allows statistical shape analysis to be performed on a dense set of 3D data points. Three-dimensional imaging technology makes it possible to capture detailed shape of the human body and opens up new opportunities for product design and testing. However, raw 3D anthropometric data is not readily usable for extracting shape information for a population. Unlike traditional anthropometric data which consist of one-dimensional measurements, 3D anthropometric data usually come from digitizing the surface of the human body and are typically represented as surface meshes. They are able to capture detailed shape information about the human body. However, due to the limitations of the optical sensors, for example, occlusion and lighting conditions, the raw 3D mesh data are noisy and incomplete. Much processing is needed before any statistical analysis can be performed. In this paper, we present geometry processing tools for preparing the raw data for shape analysis. Most of these tools were developed in computer graphics, computer vision, and pattern recognition.

The fundamental measurements of the 3D anthropometry are 3D points represented as

An effective way of establishing a correspondence among all the models is to fit a generic surface model to each scanned subject such that the key anatomical features are in correspondence. Anthropometric landmarks placed on the subjects prior to scanning can be used to guide the deformation of the generic model to the scan. This process can be formulated as a large-scale nonlinear optimization problem. Modern computer hardware and numerical algorithms allow us to solve this problem efficiently. Finally, we performed PCA on the parameterized dataset. The main mode of variation of the head and face shape was analyzed for all subjects combined and for males and females separately. Through these analyses, we showed that the space of the human head and face shape is spanned by a small number of basis vectors. Furthermore, the main mode of variations can be visualized through animating the shape changes along the PCA axes.

In this study, head scan data from the 2003 NIOSH U.S. civilian worker survey and another smaller study utilizing U.S. civilian test subjects were parameterized using geometry processing techniques developed in computer vision and computer graphics. Subsequently, PCA was performed on the registered dataset.

Initially 1177 head scans were available for processing (953 from the 2003 NIOSH survey and 224 from the ongoing temporal changes in fit laboratory study). The 2003 NIOSH nationwide survey of subjects obtained anthropometric data from several occupational backgrounds that utilized respirators on a regular basis: construction, manufacturing, firefighting, healthcare, law enforcement, and other occupational groups. These subjects were divided into three age strata (18–29, 30–44, and 45–65), two gender strata, and four racial/ethnic group strata [Caucasian, African-Americans, Hispanic, and Other (mainly Asian)] (

The second dataset of head scans were from an ongoing study to assess temporal changes in facepiece fit. Test subjects performed fit testing at six month intervals to investigate factors that affect changes in respirator fit over time (e.g., changes in weight). Head scans were taken using a 3dMDcranial5 System™ (3dMD LLC, Atlanta, GA) which utilizes five banks of high speed digital cameras to capture a 360° image of the subject's head. Head scan data from the two datasets were combined to create one dataset. For the combined dataset, six subjects were excluded due to lack of demographic information and two subjects were excluded due to invalid scans. Thus, the final dataset consisted of 1169 scans (1177 – 8 = 1169).

Polyworks version 10.1.6 (InnovMetric Software, Inc., Québec QC, Canada) was used to perform additional processing and measurements of all scanned images. Polyworks permits the user to create various features such as points and distances. Points were placed manually on each individual scan in the same locations as the labeled landmarks and linear distances were defined by those technician-defined points. Since the landmark data were prepared using a separate software package, they were given in a different coordinate frame than those of the 3D scans. A transformation had to be applied to the scans to align them with the landmark points. Before alignment was done to a 3D scan, a scaling factor of 0.001 was applied to transform it from microns to millimeters. Observations and experiments showed that the 3D scans needed to be first rotated around the

The goal of data parameterization is to establish a correspondence among the models. An early attempt to solve this problem adopts a volumetric approach (

A better approach is to fit a generic mesh model to each data scan (

Deforming a generic mesh smoothly to a data scan can be formulated as an optimization problem. Here, the variables that require solving are the

A typical scan consists of 100,000–300,000 points. Thus, our optimization problem involves the solving of a large number of variables. As the problem is nonlinear, it is difficult to find stable solutions and the algorithm tends to be stuck in the local minima. Allen et al. suggested a multi-resolution approach where low resolution meshes are deformed before the high-resolution meshes (

As can be seen in

Previously, Luximon et al. feared using the method of fitting a template (generic) model to individual 3D scans, as information on the accuracy of the model and methods was not then in existence (

Having established the correspondence among all the models, we can perform statistical shape analysis. At this point, we have a set of parameterized models, each containing the same number of points and the same mesh topology. The variables on which we perform statistics are the coordinates of the vertices on the meshes. A shape vector can be formed for each model and the mean vector and the covariance matrix can be computed. The eigenvectors of the covariance matrix form a basis of the shape space. This is the standard PCA method. It transforms the data into a new coordinate system in which the modes of variations are ordered from large to small. Three different types of statistical analyses were performed: all subjects combined, and male and female separately.

The absolute values of the eigenvalues determine the significance of the corresponding variations (principal components). Many of them are close to zero. Therefore, the space of the human head shape can be represented by a small number of principal components (PC).

One of the advantages of statistical shape analysis is that it provides intuitive visualization of the shape variation. Since we used a dense point set on the surface to perform PCA, each PC can be visualized by an animation produced by varying the coefficient of the component. We have implemented software, called Shape Analyzer, for navigating the shape space.

At any instance, the generated shape can be exported to a 3D file which can then be used in computer aided design software for developing headforms; these headforms, which can be kept up-to-date with changing population demographics and easily created with the shape analyzer software, could then be used for respirator or other safety device development. Since PCA is a linear model, shapes along a PC axis are expected to form a Gaussian distribution. Therefore, boundary shapes of a population can be generated. Normality tests were conducted for each PC and they were found to be normally distributed.

To present shape variations for males and females combined along each principal component, we sampled the Gaussian space at (−3_{i}_{i}_{i}_{i}_{i}

For the male and female combined dataset, the first five PCs account for 70% of the total sample variation (variability in the first five components are 35%, 15%, 9%, 6% and 5% respectively). The first PC indicates the overall size is an important component of facial variability. The second PC accounts for long and narrow or short and wide faces. For PC3, the variation is between deep/narrow head and shallow/wide head. Protruding chin versus prominent back of head can be described by PC4, and PC5 represents variations between a head with protruding lip and a head with prominent back of head.

For the male dataset, the first five PCs account for 68% of the dataset variation (variability for the first five components are 30%, 17%, 9%, 7% and 6% respectively). Variations on the first three PCs are similar as those observed for the combined dataset. PC4 depicts shape variations between a head with protruding lip and one with prominent back of head. PC5 represents variations between a head with protruding chin and a head with protruding nose and forehead.

For the female dataset, the first five PCs account for 65% of the dataset variation (variability for the first five components are 33%, 12%, 10%, 6% and 5% respectively). The first PC represents a variation between a small/narrow head and a large/wide head and PC2 indicates variations between deep/narrow and shallow/wide. A variation between long/shallow and short/deep is observed on PC3. The variation on PC4 is similar to that on PC5 in the male dataset. PC5 in the female dataset depicts a variation between a head with protruding and wide lower face and one with prominent back of head and narrow lower face.

In PCA, eigenvectors are computed to form a shape space. As a result, all the shape vectors comprised of

A plot onto the first several PCs shows differences between subjects. For the first ten PCs, we found PC1, PC2, and PC4 to be best in distinguishing the subjects. The average of mapped weights was computed and the nearest neighbor from the subjects was found. With the two averages, a straight line was drawn, the rest of the subjects mapped onto the line, and the two subjects being furthest from the averages were found.

The first ten PCs were studied and the subjects plotted from four different ethnic groups. The finding is that subjects in the “Black” group are distinguishable from the other three groups using PC3 and PC4. It is difficult to distinguish among subjects from the “White”, “Hispanic” and “Others” groups using any of the first ten components. Therefore, a scatter plot was made in

Continuing research in head shape analysis is beneficial for constructing headforms that take into account the facial form (size and shape). Such headforms could be incorporated into respirator research, certification standards and product design in efforts to create better fitting and more comfortable respirators. Standard headform specifications exist globally and are used to test the efficacy of various types of personal protective equipment (PPE); however, specific designs vary depending on the PPE being tested and applicable test standards.

Recently, Zhuang et al. developed five digital 3D headforms using the 2003 NIOSH survey data by computing mean facial dimensions to target the ideal facial dimensions for each size category; five scans of test subjects in each category were then chosen and averaged to construct a representative headform for each size category (

Point-cloud analysis of 3D head scan data can reveal detailed shape variation among populations and, in turn, can be utilized for the design of better fitting respirators. Using a large digital dataset taken from two recent 3D anthropometric surveys (SizeChina and CAESAR) (

Luximon et al. also used the data from the SizeChina survey in their research. In an effort to create an accurate 3D head-and-face model representative of the current Chinese population, the researchers employed the usage of virtual landmarks in addition to physical landmarks to better obtain the entirety of the head shapes of each scanned individual (

The newly implemented Shape Analyzer software is of crucial importance because it not only makes it possible to create 3D digital headforms of virtually any head-and-face size and shape, but it also makes it simple to then use computer aided design software to evaluate and test new product designs. Such headforms are significant because they can be newly and quickly constructed as the demographics of the US population – or even a specific sampling population – change. These new 3D digital headforms can also be used to create physical headforms which could help manufactures develop and test their new respirators, ensuring better protection and more comfort.

Of further note, Zhuang et al. used varying methods to reach the same conclusions described here (

This study found that the first 33 PCs explain more than 90% of the shape variability. Similarly, a landmark based analysis using 26 facial landmarks found that the first 27 PCs, as a group, account for 90% of the total sample variation (

Comparisons between the previous studies (

Previous studies have identified areas of the face where leakage occurs for respirators. Roberge et al. observed that the nasal region and malar (cheekbone) regions accounted for 71% of identified leak sites in subjects wearing filtering facepiece respirators (FFRs) (

Geometry processing techniques developed in computer vision and computer graphics were used to prepare the data such that they correspond to each other point wise. Subsequently, PCA was conducted on the registered dataset. The PCA provides a compact description of human shape variability. Less than 50 components usually explain more than 90% of the variability. Furthermore, the main mode of variation could be visualized through animating the shape changes along the PCA axes with computer software in executable form for Windows XP. The results from this study, in turn, could feed back into respirator design to achieve safer, more efficient product style and sizing. Future study is needed to determine the overall utility of the point cloud-based approach for the quantification of facial morphology variation and its relationship to respirator performance.

Disclaimer

The findings and conclusions in this report are those of the authors and do not necessarily represent the views of the National Institute for Occupational Safety and Health. Mention of commercial product or trade name does not constitute endorsement by the National Institute for Occupational Safety and Health.

From left to right: the generic model (top) and the original scan (bottom), the deformed shape of the generic model using radial basis function, and the generic model after fine fitting to the original scan.

From left to right: the original scan, fine fitting result and fine fitted model textured with color-coded errors. The values on the scale are in millimeters.

Plot of percentiles for male and female combined on variations represented by principal components.

Interface of the shape analyzer.

Shape variations for males and females combined along the first five principal components (from left to right are reconstructed shapes at −3_{i}_{i}_{i}_{i}

Shape variations for males along the first five principal components (from left to right are reconstructed shapes at −3_{i}_{i}_{i}_{i}

Shape variations for females along the first five principal components (from left to right are reconstructed shapes at −3_{i}_{i}_{i}_{i}

Front and side views of average head models. From left to right: average male and female combined head, average male head, and average female head.

Scatter plot of mapped weights on the first and second principal components by gender, where B and C denote the nearest neighbors to the average male and female, A is the furthest to B in males and D is the furthest to C in females.

Scatter plot of mapped weights on the third and fourth principal components, where B and C are the neighbors to averages of “White” and “Black”, A denotes the furthest subject in “White” to B and D represents the furthest one in “Black” to C.

The number of head scans in different genders.

Gender | Male | Female |
---|---|---|

Number of scans | 784 | 385 |

The number of head scans in different ethnicity groups.

Ethnicity | White | Black | Hispanic | Others |
---|---|---|---|---|

Number of scans | 667 | 260 | 175 | 67 |