Human Anthropometric Parameters Estimation Using Video Based Techniques

IVO STANCIC, TAMARA SUPUK, MOJMIL CECIC Laboratory for Biomechanics and Automatic Control

Faculty of Electrical Engineering, Mechanical Engineering and Naval Architecture, University of Split Rudjera Boskovica b.b. 21000 Split

Croatia istancic@fesb.hr, tada@fesb.hr, mcecic@fesb.hr, http://labacs.fesb.hr

Abstract: - In this paper simple and fast method for human anthropometric parameters estimation was proposed and evaluated. Anthropometric data obtained by this method are human segments lengths and masses, which could be primarily used for biomechanics applications, but could be adapted to other fields, as well. Measurement system designed for the purposes of this work consist of a single digital camera, PC and requires simple experimental set–up. We developed 3D model of the human body by using data obtained from 2D subject’s image and ellipses as basic objects. Advantages of the proposed method are contactless measurements on the subject which result in minimizing examiners physical intervention and considerable reduction of time required for measurements. Therefore, measurements can be conducted on a large number of subjects, in a short period of time. Proposed method comes with some disadvantages, due which it could be rather used as an estimation tool than as a highly-precise measurement procedure of anthropometric parameters, which is usually performed on cadavers. Key-Words: - biomechanics, human anthropometry, computer vision, image processing 1 Introduction The measurement of the human anthropometry is complicated and time consuming procedure. In order to obtain highly precise results, each measurement has to be done by the experienced examiner using various methods and techniques for each parameter [1]. Some anthropometric measurements were impossible or impractical to be done on the living human (e.g. distribution of the masses of each body segment) or result with irradiation of human body [1,2]. Some of these measurements were done on cadavers, which result in errors, due to different physical characteristics of living and deceased human body. Generalized anthropometry tables were built for different races, sexes, and even body types [1,3]. Using those tables, and with knowledge of basic information about subject’s body (weight and height) it is possible to calculate anthropometry parameters for each person. Results are instantly obtained, but are unsatisfactory for biomechanics applications, due to tables that are in generalized form and not prepared for each individual. In some applications, like precise garment manufacturing it is required to obtain correct anthropometry data of every individual, measurement are usually done using highly sophisticated body scanners [4,5,6,7,8]. Body scanners are today too expensive for most of biomechanics laboratories, and are only available to the large companies and organizations. Thus, development of system that could estimate anthropometry parameters for

each individual, with its complexity and cost scaled somewhere between generalized anthropometry tables and full body scanners makes sense. Such system is proposed and evaluated in this paper. 2 Human anthropometric parameters A wide variety of physical measurements are required to describe characteristic of the human body for the needs of today technological developments and applications, especially man-machine interfaces, workspace and individual garment design [9,10]. Most of these needs are satisfied by the simplest linear and volumetric measurements. However, biomechanics analysis of the human movement requires kinetic measurement as well; knowledge of segments masses, moments of inertia and their positions in human body are mandatory. These measurements are, as mentioned in the previous section, complicated and require vast knowledge of a human anatomy (joints center of rotation, muscle placements, etc). In biomechanics, human body is described as a body composed of rigid segments connected in the joints. Rigid body is assumed to be object that deformations of shape are consider negligible for current application and can be ignored [9,10]. Physical characteristics of body segments are measured, statistically arranged and filled in special prepared anthropometry tables. Number of body segments in available anthropometry tables vary between authors and applications [1,3]. In this work, we divided human body

Proceedings of the 10th WSEAS International Conference on AUTOMATION & INFORMATION

ISSN: 1790-5117 169 ISBN: 978-960-474-064-2

in to the 14 segments, as follow: left and right hand, forearm, upper arm, foot, shank, tight; trunk and head + neck segment, as depicted in Fig. 1.

Fig.1, Human body divided into the 14 segments

3 Measurement and methods Anthropometric measurement method proposed in this work is planned to be fast, simple, and adequate for examiners with no experience in traditional anthropometric measurements. Method gives fast and rough analysis of a human body anthropometry while keeping acceptable accuracy. Using digital camera instead of mechanical measuring devices, and computer program that takes care of all calculations and analysis, measurement process is simplified and semi-automated. This method requires minimal physical and psychical involvement of the examiner. 3.1 Measurement equipment For correct execution of the measurements, simple setup was required. One vertical surface has to be covered with the uniform single color, in order to create background for Chroma key technique [11,12]. For this work, single peace of green fabric with dimensions of 2.5m x 2m was used, large enough to completely cover camera’s filed of view, as shown in Fig. 2.b) Camera used in this work was Canon Power shoot A560 digital camera with 7.1 Mpixels sensor and capability of recording videos with VGA (640 x 480 pixels) resolution at 30 frames per second. Hama Gamma 74 camera tripod was used for an optimal camera placement and orientation in the setup. For image processing and data analysis Personal computer was used with the following configuration: Core 2 duo processor @ 2.66Ghz, 2Gb RAM Microsoft Windows XP SP3 operating system. Image processing and data analysis were done using MathWorks Matlab 2006 software package. Measurement setup requires that complete cameras sight has to be cleared of any

obstacles. Another mandatory requirement was adequate lighting conditions, in order to achieve best possible image of measured subject, without sharp shadows on the background or uneven lighting of a subject. Best result are obtained using professional studio lighting equipment, but even daylight lighting with help of few extra lights or small reflectors gave satisfactory results. 3.2 Measurement setup As mentioned in the previous section, one vertical surface was covered with the uniform green fabric using a specially prepared mechanical construction, as shown in Fig. 2. Camera was placed on a tripod about tree meters from the background, vertically oriented in order to optimize utilization of a camera sensor. Calibration board was placed on the ground in the vertical position 50cm from the background, as shown in Fig. 2.a). Process of calibration is explained in section 3.3. After calibration process, subject was placed between camera and background, 50cm close to the background in the same plane as the calibration board, which is important because once camera is calibrated and results are valid only in that plane.

a)

b)

Fig.2, Measurement setup: a) with the calibration board, b) with the subject

Proceedings of the 10th WSEAS International Conference on AUTOMATION & INFORMATION

ISSN: 1790-5117 170 ISBN: 978-960-474-064-2

3.3 Measurement process Four different measurements were done on a single subject, first three for hand, forearm and upper arm segment length estimation and one for a body segments volume and mass estimation. For the purpose of testing and method validation, only anthropometric parameters of hand, forearm, and upper arm were measured. In the first three measurements, video was required for process of determination segment lengths, as described in section 4.1. Similarly, for the last measurement of body anthropometry only one single high resolution image was required. Pose and movements of the subject during measurements are shown in Fig. 3 and are described as follows: in arm length measurement subject was instructed to standing still, oriented perpendicular to the camera’s image plane, that is facing his sagittal plane.

Fig.3, Subject movements during arm length measurement

Fig.4, Subject movements during forearm length measurement

Fig.4, Subject movements during hand length measurement

Video of the subject was recorded while he was flexing and extending his arm in a shoulder. Similar procedure was followed for forearm and palm anthropometry measurements. In the same pose, with his arm horizontally extended subject is flexing and extending hand or forehand in joint while video of that movement was recorded, as shown in Fig. 4. and 5. In the last measurement, subject was facing camera with his frontal plane, his legs and arms were slightly extended. Subject was instructed to copy the pose of Leonardo’s Vitruvian Man, as shown in Fig.1. In this pose, limbs are better distinguished from the trunk by image processing algorithm.

3.4 Camera calibration In order to obtain the most accurate measurement as possible, camera had to be calibrated. Calibration process was done by using simple calibration board with the size of 60 cm width and 80 cm height. Board was placed in vertical position parallel with the background plane, as shown in Fig. 2.a) With camera set up and fixed in desired position, one image of background with calibration board was taken. Goal of this calibration process is to eliminate geometric distortion due to the imperfect camera setup (e.g. camera plane not parallel with the measurement or background plane, small inclination of the camera) and to transform image pixel information to the image described by metric system. Calibration process was done using technique based on Camera Homography and Projective transformation, where original image with position of minimum of four selected points (edges of calibration board) is compared with the same points in the virtual image of an object, with its size in centimeters [13,14]. Result of the calibration is transformation matrix, which transforms all pixels from original image to calibrated transformed

Proceedings of the 10th WSEAS International Conference on AUTOMATION & INFORMATION

ISSN: 1790-5117 171 ISBN: 978-960-474-064-2

image, where every pixel is equal of 0.2mm in length. Calibration matrix obtained this way could be used for calibrating every image that was taken with the same camera setup. 4 Anthropometry estimation method Measurement videos and images were obtained from the digital camera. By using video editing software, video sequences of measurement were extracted and then analyzed frame by frame. Goal of image processing used in this work is to extract subject from the background, detect moving body parts in video sequence and finally estimate their characteristics. For extraction of the subject from background Chroma key technique was used. Chroma key technique is well known and used for years in the 7.th art [11]. Original image or frame was represented in RGB color space by array of pixels, where each pixel, green, red, and blue respectively, can take one value in range [0-255]. Image was then transformed from RGB to HSV color space, which is more powerful in describing variations of a single color (hue), [12]. Values of backgrounds Hue, Saturation and Intensity were obtained experimentally, and must have range of values chosen carefully to cope with problems that could arise during measurements, like changing lighting conditions, shadows and fabric perforations. Image and frames were transformed by the calibration matrix, as explained in section 3.4. Next step was to extract subject from the image. After background removal, some noise due to imperfection of equipment and technique is still left on the image. Binary image was created from resulting image and by using the operations of binary image dilatation and erosion, small imperfections from extracted subject were repaired and some noise was removed. Largest object in resulting image represent the subject, while other small objects are considered to be noise. With selecting only the largest object in image, subject was successfully extracted from the background and prepared for the further analysis. 4.1 Determination of segments length For process of determination of segments length, only moving body parts had to be extracted and analyzed from video, while still body parts were added to the background. Result is a series of binary images of subject’s moving body parts during video sequence. During the rotation movement (e.g. arm flexing and extending in shoulder), the point that is moving with the highest velocity is considered to be the endpoint of a moving segment. By knowing the endpoint of that moving segment in all frames, it is possible to precisely calculate center of rotation using circle fitting technique, Fig. 6, [13,14]. Estimated center of rotation is in fact estimation of joint centre, while radius of fitted circle is

length of the moving object. By repeating this process for each body segment measured in this work, all segments lengths are finally estimated.

Fig.6, Center of rotation (joint center) of a moving body

segment 4.2 Dividing of subject’s body into the segments Division of subject’s body in to the segments starts from image of subject in Vitruvian Man pose as shown in Fig.1. After body extraction from the background is completed, five most distant points from the estimated center of body have to be found. These points represent tips of hands, foots and vertex. In order to divide subject’s body in segments, we developed special image processing method has to be developed, and explained as follow. Subject’s body was scanned using image processing algorithm, starting from the tip of body limbs previously obtained. Each limb was scanned with lines perpendicular to the limb’s longitudinal axis. After length of segment was reached, scanned body segment was extracted and removed from the original binary image. The process of ‘’limb scanning’’ continues until trunk is reached, and after all limbs were processed including head and neck, remaining object from original image is subject’s trunk, which is assumed to be rigid body segment undivided into smaller segments.

4.3 Estimating body segments volume and mass To calculate body’s segment mass, segments volumes had to be found first. In this work, human body is assumed to be of homogenous density, with constant segment density factor [1,3]. Mass of each segment was calculated by multiplying estimated volume with its density factor. It is hardly possible to precisely calculate volume of body segments based only on a single image, so we had created body model which needs only 2D image information. In all 3D body models, body is represented by series of geometric objects connected together, crating one large body. In this work we modeled body segments with a series of ellipses of

Proceedings of the 10th WSEAS International Conference on AUTOMATION & INFORMATION

ISSN: 1790-5117 172 ISBN: 978-960-474-064-2

different size but constant ratio of longest and shortest side [8, 15].

Fig.7, Hand and forearm modeled by series of ellipses

Ratios were determinate experimentally for each body segment. In our model, only one requirement had to be followed, cross section of body’s segment must be equal of 2D image representation of that body. For each segment that was extracted by method described in previous section, its binary image was rotated so the longest dimension is parallel to vertical axis. Binary image was than scanned using the horizontal lines, where length of line between objects edges is equal to width of the body segment in a given point. Line in 2D image is modeled as ellipse in 3D using equation (1), where x is distance between point in scanning line and segments vertical axis and value y is calculated.

12

2

2

2

=+by

ax

(1)

Fig.7, Ellipse as fundamental modeling object of our

human 3D model

Resulting object is half-ellipse, while the full ellipse is obtained by simple mirroring half ellipse in image plane. Finally, volume of body segment was calculated by summing all volumes defined by ellipses edges.

5 Results and discussion In this section, results of an anthropometric measurement and estimation by three different methods will be shown compared, and discussed. Body segments lengths were measured using tape meter, where definitions which are followed of the segments beginnings and endings are found in the literature [16]. That first measurement is considered to be most accurate, and considered to be referent in this work. Results of second method were obtained using generalized anthropometry tables, where subject’s height was multiplied with the segment’s length coefficients found in anthropometry tables [1,3]. Results of measurements and estimation by first two methods were then compared with a third method, Computer Vision method described in this work. All measurements were preformed on two males aged 24 (Subject 1) and 27 (Subject 2) and one female aged 24 (Subject 3) years. All subjects were Caucasian. Table 1, Comparison of the results obtained by different methods

Length of the segment Body

segemnts Measured by meter

Anthrop.tables

Comp. vision

Hand 20.0 cm 19.5 cm 19.5cmForearm 28.0 cm 27.5 cm 26.5cmUpper arm 30.0 cm 34.0 cm 30.5cm

Sub. 1

Whole arm 78.0 cm 81.0 cm 76.5cmHand 18.5 cm 19.0 cm 17.5cmForearm 25.0 cm 25.5 cm 25.5cmUpper arm 29.0 cm 32.5 cm 27.0cm

Sub. 2

Whole arm 72.5 cm 77.0 cm 70.0cmHand 18.0 cm 19.5 cm 18.5cmForearm 28.0 cm 26.5 cm 24,5cmUpper arm 29.0 cm 33.5 cm 33.5cm

Sub. 3

Whole arm 75.0 cm 79.5 cm 76.5cm Results from Table 1 shows that measurements with the computer vision system are in most cases more accurate than simple estimation of the segments length using anthropometry tables. In Table 1, results that are closer to referent value are grayed, if both results are equally close to the referent value, both results are grayed. Experience during measurement show that small anomality while making predescribed movements of arm segments, small flexion in joint that is assumed to be

Proceedings of the 10th WSEAS International Conference on AUTOMATION & INFORMATION

ISSN: 1790-5117 173 ISBN: 978-960-474-064-2

fixed, could affect measurement results, and thus show inaccurate segment lengths. This problem could be assisted by using one more camera placed on the ceiling, which would be used to adjust results by detecting flexion in all joints. Accurate segments masses could not be measured for the needs of this work, thus anthropometry tables are considered to be referent for the second comparison. Body segment’s masses estimated using our method were then compared by Anthropometry tables data and filled in Table 2. Table 2, Comparison of mass coefficients obtained by Anthropometry tables [3], and Computer Vision system

masssubjectmasssegment

Antroph. table [3] Comp. Vision Hand 0.006 0.006 Forearm 0.016 0.017 Upper arm 0.028 0.030 Foot 0.061 0.010 Leg 0.046 0.048 Tight 0.100 0.119 Head + neck 0.081 0.063 Trunk 0.497 0.495

Comparison of the results from Table 2 validates most of our segment’s mass estimations. Foot and Head + Neck body segment mass estimation is not matching referent anthropometry data. Problem is in correct estimation of segments volumes from single image, and some improvements of our method should be considered. 6 Conclusion Proposed method of a fast estimation of the human anthropometry showed to be promising, especially in situations where only fast and estimation is required on large number of subjects. As previously assumed, manual measurement by mechanical devices of subject’s anthropometry is more accurate (and in this work was considered to be referent), but requires a lot of valuable time. As compared with method that uses generalized anthropometry tables, our results are personalized for each subject and thus showed to be better for biomechanics application. Proposed upgrading of this method would be anthropometry estimation by multiple digital cameras. Improvements of such system would give improved and more accurate 3D model of human body, by which the values of measured anthropometric parameters would be comparable by experimental full body scanners.

References: [1] R. Drills , R. Continti, M. Bluestein, Body Segment

Parameters; A survey of Measurement Techniques, Artificial Limbs, Vol 8, Num 1, 1964, pp. 44-66.

[2] V. Zatsiorsky V.Seluyanov, The mass and inertia characteristics of the main segments of the human body. Biomechanics, Vol.VIII-B, 1983, pp.1152-1159.

[3] David A. Winter, Biomechanics and Motor Control of Human Movement, Wiley; 3 edition, 2004.

[4] Jun-Ming Lu Mao-Jiun J. Wang, Automated anthropometric data collection using 3D whole body scanners, Expert Systems with Applications, Pergamon Press, Inc, July 2008 ,Volume 35 Issue 1-2, pp. 407-414.

[5] Fabio Remondino, 3D Reconstruction of Static Human Body with a Digital Camera, Proc. of Videometrics VII - El-Hakim, Gruen, Walton (Eds), SPIE Electronic Imaging, Vol. 5013, USA 2003, pp. 38-45.

[6] Fabio Remondino, Andreas Roditakis, Human Figure Reconstruction and Modeling from Single Image or Monocular Video Sequence, 4th International Conference on 3D Digital Imaging and Modeling (3DIM), 2003, Banff, Canada.

[7] Pierre Meunier and Shi Yin, Performance of a 2D image-based anthropometric measurement and clothing sizing system, Applied Ergonomics, Vol. 31, Issue 5, 2000, pp. 445-451.

[8] J Starck, A Hilton and J. Illingworth, Human Shape Estimation in a Multi-Camera Studio, In Proceedings of British Machine Vision Conference, Vol 2, 2001, pp. 573-582.

[9] Duane Knudson, Fundamentals of Biomechanics, Springer, 2007.

[10] NASA, Man-Systems Integration Standards NASA-STD-3000, Volume I, Revision B, 1995.

[11] Ron Brinkmann, The Art and Science of Digital Compositing, Morgan Kaufmann, 1999.

[12] G. Shapiro and George C. Stockman, Computer vision. Prentice-Hall 201, USA 2001.

[13] Richard Hartley, Multiple View Geometry in Computer Vision – second edition, Cambridge university press 2003.

[14] Emanuele Trucco, Alessandro Verri, Introductory Techniques for 3-D Computer Vision, Prentice Hall, 1998.

[15] T. Jurčević Luluć, O.Muftić and A. Sušić, Design of the 3D model of human body oriented to biomechanical applications, In Proceedings of the International design conference - Design 2004, Croatia 2004.

[16] Paolo de Leva, Adjustments to Zatsiorsky-Seluanov's Segment Inertia Parameters, J. of Biomechanics, Vol 29 No9, 1996, pp. 1223-1230.

Proceedings of the 10th WSEAS International Conference on AUTOMATION & INFORMATION

ISSN: 1790-5117 174 ISBN: 978-960-474-064-2