*Corresponding author. Tel.: #1-416-635-2093; fax #1-416-635-2013.

E-mail address: pmeunier@dciem.dnd.ca (P. Meunier).

Applied Ergonomics 31 (2000) 445}451

Performance of a 2D image-based anthropometric measurementand clothing sizing system

Pierre Meunier!,*, Shi Yin"

!Systems Modelling Group, Defence and Civil Institute of Environmental Medicine, 1133 Sheppard Ave West, P.O. Box 2000,Toronto, Ont., Canada M3M 3B9

"Image and Vision Systems, 1202-30 Godstone, Toronto, Ont., Canada M2J 3C6

Received 9 September 1999; accepted 31 March 2000

Abstract

Two-dimensional, image-based anthropometric measurement systems o!er an interesting alternative to traditional and three-dimensional methods in applications such as clothing sizing. These automated systems are attractive because of their low cost and thespeed with which they can measure size and determine the best-"tting garment. Although these systems have appeal in this type ofapplication, not much is known about the accuracy and precision of the measurements they take. In this paper, the performance of onesuch system was assessed. The accuracy of the system was analyzed using a database of 349 subjects (male and female) who were alsomeasured with traditional anthropometric tools and techniques, and the precision was estimated through repeated measurements ofboth a plastic mannequin and a human subject. The results of the system were compared with those of trained anthropometrists, andput in perspective relative to clothing sizing requirements and short-term body changes. It was concluded that image-based systemsare capable of providing anthropometric measurements that are quite comparable to traditional measurement methods (performed byskilled measurers), both in terms of accuracy and repeatability. Crown Copyright ( 2000 Published by Elsevier Science Ltd. Allrights reserved.

Keywords: Anthropometry; 2D-measurement system; Automated clothing sizing

1. Introduction

In spite of highly standardised protocols designed tomaximize the degree of repeatability and accuracy ofmeasurements, anthropometric data are not always asreliable as they appear. Many factors come into playduring the measurement of human subjects, which canresult in the appearance of numerous sources of error.Some of the important sources include posture, identi-"cation of landmarks, instrument position and orienta-tion, and pressure exerted by the measuring instrument(Davenport et al., 1935). The di$culty in controlling allpotential sources of error is such that it has been said thattrue values are seldom measured in anthropometry(Jamison and Zegura, 1974). Accuracy and precision ofanthropometric measurements are at the mercy of themeasurers who take them. Even if measured by highly

trained observers, comparison of two populations maybe meaningless (Bennett and Osborne, 1986). In a com-parative study by Kemper and Pieters (1974), 50 boys (12and 13 years of age) were measured independently byexperienced observers in two institutes. Both teams ofobservers were trained to the same measurement tech-niques and used the same measuring instruments. In spiteof this, systematic di!erences were found in nine of the 12measurements taken. Pearson correlations between 0.872(biacromial diameter) and 0.996 (stature) were found be-tween the measurements taken by the two groups. Al-though the variable with the lowest correlation(biacromial diameter) did not present systematic errors, itsu!ered from repeatability problems (precision error).The results of these and many more studies show howdi$cult it is to measure humans, even under controlledconditions and after extensive training of the observers.

Computerized image-based systems o!er an interest-ing alternative to traditional methods of measurement,especially in applications where individuals are measuredto determine their size of clothing and equipment.They o!er rapid body measurements that are quickly

0003-6870/00/$ - see front matter Crown Copyright ( 2000 Published by Elsevier Science Ltd. All rights reserved.PII: S 0 0 0 3 - 6 8 7 0 ( 0 0 ) 0 0 0 2 3 - 5

Fig. 1. Plan-view of image capture set-up.

Fig. 2. Front and side silhouettes with landmarks.

translated into clothing size categories based on pre-determined "t criteria. Such systems are capable of over-coming some of the problems of traditional anthropo-metry, but introduce error sources of their own. Inimage-based systems, the sources of error take the formof perspective distortion, camera resolution, camera cal-ibration, landmarking error, and modelling error (sincecircumferences are not measured directly). The objectiveof this paper is to compare the accuracy and precision ofmeasurements made from two-dimensional images ofhumans with those of highly trained anthropometrists,and put the results in perspective in the context of cloth-ing and equipment sizing.

1.1. System description

The system under review is a PC-based system com-prised of two Kodak DC120 colour digital cameras(1280]960 pixels) and a blue backdrop embedded withcalibration markers (Fig. 1). The calibration of thecameras is performed using an algorithm developedby Tsai (1986). The system takes simultaneous frontand side pictures of individuals standing with their armsstraight and slightly abducted along their side . By takingboth images simultaneously, the exact posture in space iscaptured, and it is possible to recover the object dimen-sions in 3D. Fig. 2 shows the "nal result, after imageprocessing and automatic landmarking. The cross-hairsidentify the location of the landmarks.

Potential sources of error can be found at each of thesteps in the image analysis process illustrated in Fig. 3,namely in

(a) pre-processing of the front and side images,(b) calibration of the cameras,(c) segmentation of the body from the background,(d) detection of the landmarks, and(e) calculation of the anthropometric dimensions.

1.2. Theoretical assessment of error

The error of a measurement is de"ned as the di!erencebetween the measured value and the true value of theitem being measured. Errors can be catalogued as eitherrandom (precision error) or systematic (bias error). Pre-cision is de"ned as the di!erence in values obtained whenmeasuring the same object repeatedly. It has an averagevalue of zero. Accuracy is the di!erence between themeasured and true values. Bias error, which occurs in thesame way on each measurement, a!ects the accuracy ofa measurement while random error a!ects precision.

The concept of error is useful, but it implies knowledgeof the true value of what is being measured. Since anymeasurement contains error, the true value is neverknown. Although absolute error cannot be calculated, itcan be estimated. Precision error, for instance, can beestimated by taking a large number of readings on anindividual and using a statistical model to determine theexpected spread of values at a given probability level.Bias error, on the other hand, requires comparison ofmeasurements with a more accurate method/instrument.This is di$cult to do in anthropometry, given that thebest available method, i.e. traditional measurementstaken by highly trained and skilled measurers, is one thatcontains non-negligible error itself, as discussed in theintroduction.

Of the sources of error listed above (a}e), segmentationand modelling are probably the most signi"cant. A roughestimate of segmentation error can be made from a theor-

446 P. Meunier, S. Yin / Applied Ergonomics 31 (2000) 445}451

Fig. 3. Image analysis process.

etical perspective, using camera resolution as the startingpoint. Since the cameras used in this system have1280]960 pixels covering an area that is approximately2.5]1.8m at the subject, the corresponding spatial res-olution is slightly less than 2.0mm/pixel. This resolutiontends to remain constant regardless of camera position,since the camera set-up procedure requires a full view ofthe backdrop, which is achieved either by moving thecamera and/or by adjusting the zoom. Assuming a seg-mentation error of plus or minus one pixel, directmeasurements requiring two points (i.e. for breadths,depths, and heights) will #uctuate within $2mm of thetrue value (1 pixel]2mm/pixel). The maximum error,which is obtained when both points err in opposite direc-tion, would put the result within $4mm of the truevalue (2 pixels]2mm/pixel).

Circumferences cannot be measured directly usingonly front and side pictures and must therefore be cal-culated using some form of mathematical modelling. Thechoice of model depends on the cross-sectional shapebeing measured, which varies among individuals. Intwo-dimensional systems, circumference measurementerror depends on the accuracy of the model as well as onthe breadth and depth measurements used in the calcu-lation. Assuming a cylindrical object (i.e. no modellingerror) and a one-pixel error on the diameter, the resultswould be likely to #uctuate within $6mm (p]2mm) ofthe true value.

2. Methodology

2.1. Accuracy assessment

The accuracy of the image-based system was assessedby comparing image-based measurements with manual

measurements taken by anthropometrists during the1997 survey of the Canadian Land Forces (Chamberlandet al., 1998). Six dimensions were selected because of theirrelevance to clothing sizing, which is the main purpose ofthe system. These were: stature (as a general indicator oflength measurements), neck circumference, chest circum-ference, waist circumference, hip circumference, andsleeve length (spine-wrist). The de"nitions for stature,and for neck, chest, and hip circumferences were virtuallyidentical in both methods of measurement: stature wastaken at the highest point, the neck at the Adam's appleperpendicular to the long axis of the neck, chest horizon-tally at the fullest part of the breast, and hip horizontallyat the maximum protrusion of the buttock. The de"ni-tions for waist circumference were di!erent: the tradi-tional measurement of waist circumference was takenhorizontally at the level of the navel, whereas the image-based system used what is sometimes referred to as `pre-ferreda waist, i.e. where individuals wear the belt of theirtrousers. The side view in Fig. 2 shows that contrary tothe traditional de"nition, the waist landmarks are nothorizontal. The traditional spine-wrist sleeve length def-inition is somewhat incompatible with front and sideviews of a subject, since it requires a di!erent posturethan that used in the system and would require a topview of the subject. However, it is possible to obtain anequivalent measurement from measurements that areaccessible from those views, namely sleeve outseam (fromthe acromion to the wrist) and biacromial breadth.

The test sample consisted of a subset of 349 subjects(95 females and 254 males) from the survey that had beenmeasured both with traditional methods and with theimage-based system. Each subject was measured once byeach method. The image-based system was the last ofseven anthropometric measurement stations in the sur-vey (Chamberland et al., 1998), which means that both

P. Meunier, S. Yin / Applied Ergonomics 31 (2000) 445}451 447

Table 1Accuracy results (mm)

Measurement Females Males

Mean Std. dev. Correlation Mean Std. dev. Correlation

Stature: n"95 n"254Manual 1633 60 0.98 1747 64 0.982D system 1633 59 17487 63

Neck circumference: n"62! n"254Manual 329 18 0.88 395 23 0.942D system 329 16 395 22

Chest circumference: n"88! n"254Manual 956 87 0.95 1024 83 0.942D system 957 84 1024 78

Hip circumference: n"95 n"238"

Manual 1027 91 0.98 1005 72 0.942D system 1026 89 1004 68

Sleeve length: n"95 n"254Manual 799 34 0.79 876 35 0.762D system 800 27 875 26

!Some subjects were rejected due to interference of hair with landmarking."Some subjects were rejected due to clothing interference (boxer shorts).

Table 2Mannequin repeatability results (mm)

Variable Mean(n"10)

Range Std. dev. 1.96 Std.dev.

Stature 1822 2 1 1Neck circumference 360 5 2 3Hip circumference 946 12 4 7Waist circumference 856 7 2 4Chest circumference 960 10 3 6Sleeve length 829 24 8 15

Table 3Human repeatability results (mm)

Variable Mean(n"10)

Range Std. dev. 1.96Std. dev.

Stature 1817 5 2 3Neck circumference 369 6 2 4Hip circumference 978 11 4 8Waist circumference 873 15 5 10Chest circumference 964 16 6 11Sleeve length 887 36 10 20

types of measurements were performed within a period ofabout 90min, avoiding the e!ects of daily body vari-ations. t-tests were performed to compare the means ofall dimensions, and F-tests to compare the standarddeviations. Waist circumference was excluded from thiscomparison due to the di!erence in measurement de"ni-tion between the two methods.

2.2. Precision assessment

The precision of the image-based system was deter-mined by performing ten repeated measurements ona full size plastic mannequin as well as on a humansubject. All image capture and analysis sequences wereperformed in succession (every minute or so) such thatcamera calibration and lighting conditions were relative-ly constant. The mannequin was used in order to excludevariations due to breathing movement and postural dif-ferences from picture to picture. The human subject wasinstructed to stand with the arms slightly abducted alongthe side the body during picture taking, and to moveaway from the platform between measurements. Thus,contrary to the mannequin, the human precision esti-mates contain variability due to postural di!erences,breathing movement, and repositioning from one set ofimages to the other.

3. Results

3.1. Accuracy

The means and standard deviations for the subjectsmeasured manually and digitally are listed in Table 1.

t-tests did not indicate any signi"cant di!erence betweenthe manual and digital methods for either males or fe-males. Table 1 also lists the Pearson correlation coe$-cients between manual and 2D image measurements.

3.2. Precision

Table 2 summarises the results of repeated measure-ments of a plastic mannequin.

448 P. Meunier, S. Yin / Applied Ergonomics 31 (2000) 445}451

The results of repeated measurements of a humansubject are shown in Table 3.

4. Discussion

4.1. Accuracy

The results of t-tests comparing manual and image-based measurements showed no signi"cant di!erencebetween the two, indicating that the two methods pro-vide similar results. This is not surprising since the in-direct measurement models of the image-based systemwere developed and optimized using the images and datacollected in the survey. F-tests comparing the standarddeviations of the two types of measurements also showedno signi"cant di!erence, indicating that they wereequally consistent in taking those measurements.

4.2. Precision

The theoretical assessment of the random measure-ment error made earlier suggested that an error of theorder of $2 and $6mm could be expected froma one-pixel segmentation error on linear and circum-ferential measurements, respectively. As shown inTable 2, the results of repeatability tests performed on theplastic mannequin showed the actual error to be within1mm of the mean for stature, 95% of the time. The factthat this is less than what could be expected from a one-pixel segmentation error indicates that the spread ofmeasurements is probably due more to camera calib-ration or modelling #uctuations than segmentation.Where the mannequin's shape attributes were true to life(i.e. except for hinged joints, non-standard posture andunnatural shapes), reliable landmark positions were ob-tained using the automatic landmarking algorithms.Hinges at the shoulder, elbow and wrist hindered therepeatability of sleeve length measurements. Fluctuationsin this measurement in particular were unavoidable be-cause the automatic landmark detection software wasdeveloped to recognise real human shape. Circumferen-ces were found to be #uctuate within 3}6mm of themean, 95% of the time (Table 2), with neck circumferenceexhibiting the highest degree of repeatability. This isthought to be due, in part, to the special attention paid tothis measurement during development of segmentationand landmarking algorithms.

The results in Table 3 show that, for the most part,repeated measurements of a human subject exhibited thesame basic trend as for the mannequin in that the linearmeasurements were more precise than circumferences,and neck circumference was more repeatable than othercircumferences. In most cases, the human results exhib-ited more variability in measurement than in the caseof the mannequin, which was anticipated. The largest

di!erence between mannequin and human subjectmeasurements were for waist and chest circumferences.This can be partly explained by torso movement duringbreathing (due to expansion and contraction of the ribcage and abdomen) and di!erences in posture from pic-ture to picture (arm position, relaxed or tight posture).

4.3. Computer versus human measurement repeatability

The results of the repeatability study on a humansubject were compared with those of recent large-scalesurveys where accuracy and precision were monitoredthroughout. The "rst survey was conducted on CanadianLand Forces personnel in 1997 (Chamberland et al.,1998). The second survey was conducted on US Armypersonnel in 1988 (Gordon et al., 1989). Repeatedmeasurements were part of the routine during both sur-veys, although the methodology was slightly di!erent. Inthe Canadian Land Forces survey, subjects were re-measured by the same observers within minutes(10}90min of the "rst measurement (see Forest et al.,1999 for details). This can be viewed as the best-casescenario in terms of repeatability, since it is assumed thatthe same observer will measure in the same way given thesame landmarks. In the US Army survey, subjects werere-measured within minutes by a second observer. Thiscase can be viewed as the best-case scenario for repeata-bility by di!erent observers, since both observers werehighly trained on the dimensions speci"c to their measur-ing station.

The technical error of measurement (TEM), which isessentially a form of standard deviation, was used as thebasis for comparison. The TEM, or r, was calculatedusing the following equation:

r"J+ni/1A+k

j/1x2j!(1/k)A+j

k/1xjB

2

Bn(k!1)

. (1)

Fig. 4 shows the TEMs for computer measurements (ona mannequin and on a human) and compares them tothose obtained by trained human observers (single (For-est et al., 1999) and dual observer results (taken fromGordon and Bradtmiller, 1992)). Although the computermeasurements contain an additional source of error dueto automatic landmarking, the results indicate that therepeatability was similar to the single observer results forstature and neck circumference. The single observer re-sults had the lowest TEMs for all other measurements,however, followed by computer measurements on a man-nequin and on a human, followed by the measurementsmade by two observers.

The di!erences observed between mannequin andhuman repeatability results show the e!ect posture andbreathing can have on measurements. Better precisioncould be obtained by controlling these factors, if

P. Meunier, S. Yin / Applied Ergonomics 31 (2000) 445}451 449

Fig. 4. Comparison of technical error of measurement (TEM) obtainedby ICESS on a human subject and expert manual measurements.

Table 4Reliability of a 2D image-based measurement system

Reliability (%)

Stature 99.9Neck circumference 99.3Hip circumference 99.7Chest circumference 99.6Waist circumference 99.7

Table 5Manufacturing tolerances (on circumference) for Canadian Forcesdress trousers and shirts

Tolerance (mm)

Trouser Waist circumference $13Inseam $13

Shirt Neck circumference $ 3Chest circumference $13Sleeve length $13

required. It should be noted that had the survey resultsincluded landmarking error (the survey subjects had thesame landmarks during re-measurement), the resultscould have been more in favour of the computermeasurements.

4.4. Reliability

Mueller and Martorell (1988) state that two pieces ofinformation are su$cient to characterize the reliability ofan anthropometric variable: the TEM and the reliabilitycoe$cient. The reliability coe$cient (R) is an interestingmetric in that it compares the variability due to measure-ment error (r2) against the biological variability of thatdimension (sample variance s2). It is computed using thefollowing equation:

R"1!(r2/s2), (2)

where r is the technical error of measurement, s is thesample standard deviation, n is the number of subjectsand k is the number of measurements per subject.

If the measurement error is small compared to thestandard deviation of the sample then the reliability ofthat measurement will be high. Reliabilities above90}95% have been recommended for the selection ofvariables in a survey (Gordon and Bradtmiller, 1992).The reliability coe$cients obtained by image-basedmeasurement system were well above that, for the dimen-sions shown in Table 4.

4.5. Measurement accuracy requirements

The "rst part of the discussion dealt with the capabili-ties of the image-based measurement system when com-pared with skilled human measurement. But the answerto the question `how much measurement accuracy isrequired?a can only be answered in the context of theapplication. For clothing sizing, a large part of the an-swer comes from the manufacturing tolerances. Ina sense, the manufacturing tolerances represent the limitsof a trade-o! between "t of the clientele and cost of thegarment. They could be interpreted as an amount of#uctuation in garment size that has minimal impact on"t. By the same token, it could be said that given a gar-ment size, the same amount of #uctuation in bodymeasurement would also have minimal impact on the "tof a garment. An example of manufacturing tolerancesfor a military dress uniform is shown in Table 5. Usingthe above logic, Table 5 could be interpreted as anindication that a measurement error of $3 mm on neckcircumference and $13mm on the other dimensionswould have minimal impact on "t, which suggest thata measurement accuracy within those ranges should beconsidered acceptable.

From a di!erent perspective , it is also important tobalance the accuracy against short-term body variations.These variations, which occur naturally, must be accom-modated by the clothing regardless of their magnitude inorder for the clothing to be acceptable. Several bodydimensions can change substantially over short periods.Stature, for instance, has been known to change by30}50mm in a day depending on the amount of standing,walking and carrying done (NASA, 1978). Davenportet al. (1935) conducted experiments where repeatedmeasurements of one subject were made at various timesof day over a number of days by the same observer. Theresults (Table 6) show that measurements varied signi"-cantly, most notably for waist circumference, where 95%of the measurements were within $21mm of the mean.It could be argued that the clothing worn on a day-to-day basis commonly accommodates the amount of vari-ation reported by Davenport et al., and that those #uctu-ations should give an indication of the degree of accuracyrequired. Thus, it could be concluded that the highest

450 P. Meunier, S. Yin / Applied Ergonomics 31 (2000) 445}451

Table 6Results of repeated measurements of a subject at various times of dayover several days by one observer (Davenport et al., 1935)

1.96 S.D. (mm)

Waist circumference 21Chest circumference 15Neck circumference 5

measurement accuracy would be required for neck cir-cumference, i.e. 95% of the measurements should bewithin the range observed by Davenport, i.e. $5mm.

5. Conclusions

The results of this analysis showed that image-basedsystems can provide anthropometric measurements thatare quite comparable to traditional measurementmethods performed by skilled anthropometrists, both interms of accuracy and repeatability. The quality of theresults depends, in large part, on the dependability of theautomatic landmarking algorithms and the correct mod-elling of indirect measurements. Once that is achieved,however, this type of system can provide uniformmeasurement of a population regardless of where, whenor by whom, it is operated.

While a high degree of accuracy and precision in an-thropometric measurements is always desirable, it is notalways necessary. The requirement for accuracy shouldbe established on the basis of the application. Short-termbody changes, clothing design, "t, and manufacturingtolerances were used as guides to estimate this require-ment for the clothing sizing application. It was arguedfrom these data that a body measurement system shouldbe capable of measuring neck circumference within

$5 mm in order to be e!ective, and all other dimensionswithin $15 mm. From the accuracy and precision ana-lyses, it was found that the system under evaluation wascapable of this performance.

References

Bennett, K., Osborne, R., 1986. Interobserver measurement reliabilityin anthropometry. Human Biology 58 (5), 751}759.

Chamberland, A., Carrier, R., Forest, F., Hachez, G., 1998. Anthropo-metric survey of the Land Forces (LF97) (Contractor report 98-CR-15). Defence and Civil Institute of Environmental Medicine,Toronto, Ontario.

Davenport, C., Steggerda, M., Drager, W., 1935. Critical examination ofphysical anthropometry on the living. Proceedings of the AmericanAcademy of Arts and Science Boston 69 (7), 265}285.

Forest, F., Chamberland, A., Billette, J., Meunier, P., 1999. Anthropo-metric survey of the Land Forces: measurement error analysis.DCIEM No. CR 1999-041, DCIEM, Toronto.

Gordon, C.C., Bradtmiller, B., 1992. Interobserver error in a large scaleanthropometric survey. American Journal of Human Biology 4,253}263.

Gordon, C.C., Churchill, T., Clauser, C.E., Bradtmiller, B., McConville,J.T., Tebbetts , I., Walker , R.A., 1989. 1988 anthropometric surveyof US army personnel: methods and summary statistics.(NATICK/TR-89/044). US Army Natick Research, Development,and Engineering Center, Natick, MA.

Jamison, P., Zegura, S., 1974. A univariate and multivariate examina-tion of measurement error in anthropometry. American Journal ofPhysical Anthropology 40, 197}204.

Kemper, H., Pieters, J., 1974. Comparative study of anthropometricmeasurements of the same subjects in two di!erent institutes.American Journal of Physical Anthropology 40, 341}344.

Mueller, W., Martorell, R., 1988. Reliability and accuracy of measure-ment. In: Lohman, Rocha, Mortorell (Eds), Anthropometric Stan-dardization Reference Manual. Human Kinetics, Champaign, IL,USA, pp. 83}86.

NASA, 1978. Anthropometric Source Book, Vol. I: Anthropometry forDesigners (NASA reference publication 1024). NASA, Washington,DC.

Tsai, R., 1986. An e$cient and accurate camera calibration techniquefor 3D machine vision. IEEE Conference on Computer Vision andPattern Recognition, pp. 364}374.

P. Meunier, S. Yin / Applied Ergonomics 31 (2000) 445}451 451