Measurement of Human Body Segment Properties Using Low-Cost RGB-D Cameras

Abstract

An open question for the biomechanical research community is accurate estimation of the volume and mass of each body segment of the human body, especially when indirect measurements are based on biomechanical modeling. Traditional methods involve the adoption of anthropometric tables, which describe only the average human shape, or manual measurements, which are time-consuming and depend on the operator. We propose a novel method based on the acquisition of a 3D scan of a subject's body, which is obtained using a consumer-end RGB-D camera. The body segments' separation is obtained by combining the body skeleton estimation of BlazePose with a biomechanical-coherent skeletal model, which is defined according to the literature. The volume of each body segment is computed using a 3D Monte Carlo procedure. Results were compared with manual measurement by experts, anthropometric tables, and a model leveraging truncated cone approximations, showing good adherence to reference data with minimal differences (ranging from +0.5 to -1.0 dm³ for the upper limbs, -0.1 to -4.2 dm³ for the thighs, and -0.4 to -2.3 dm³ for the shanks). In addition, we propose a novel indicator based on the computation of equivalent diameters for each body segment, highlighting the importance of gender-specific biomechanical models to account for the chest and pelvis areas of female subjects.

Keywords: Kinect Azure; anthropometry; biomechanics; body segment parameters; body volume estimation; measurement science.

Figure 1

Scheme of the complete processing procedure with example results taken from one subject’s data.

Figure 2

(a) Scheme of the KPs considered for this work. Red KPs: computed by BlazePose; yellow KPs: calculated as midpoints of shoulders and hips segments; and black KPs: obtained by applying the biomechanical model to the original BlazePose KPs. (b) An example of vectors that were computed by the biomechanical model.

Figure 3

Examples of the filtering process. (a) Original point cloud, $P{C}_{raw}$. (b) Result of the coarse filtering process, $P{C}_{coarse,0}$. (c) Result of the fine filtering process, $P{C}_{fine}$.

Figure 4

Examples of the orientation correction of the fine filtering process. (a) Alignment of the bed’s normal to the Z reference axis. (b) Alignment of the bed’s Monte Carlo approximation principal components to the X and Y reference axes.

Figure 5

Example of some EPs drawn on a subject’s point cloud projected onto a 2D XY plane. Each EP is shown in black, and the points belonging to them are highlighted in different colors. Vectors used to create the remaining EPs are shown in red.

Figure 6

Examples of the body segment separation procedure. (a) $P{C}_{fine}$ of a subject on top, of which the 19 KPs are drawn in green. (b) Separation of the BSs of the subject highlighted in different colors. (c) Monte Carlo approximation of the BSs, which is necessary to estimate their volume.

Figure 7

(a) Point cloud filling example of the shoulder’s BS. Red: original points; green: artificial points created to fill vertical gaps around pointy areas; and cyan: artificial points created to fill the bottom part of the BS. (b) Alpha shape computed on the shoulder’s BS. (c) Monte Carlo points of the shoulder’s BS belonging to its alpha shape.

Figure 8

Images of the subjects with section lines that were directly drawn by an expert on the skin using a pen for the female subject and a tape for the male subject (which is highlighted in the picture). Vectors computed from the KPs were superimposed on the image to highlight any discrepancy. (a) Male subject, (b) female subject.

Figure 9

Boxplot of the resulting $\Delta V$ values. Comparison was only made for the limbs. Blue boxes refer to male data and pink boxes to female data. (a) Difference between the scanner volume and the one obtained from the literature reference, $\Delta V_{scanner-ref,BS}$. (b) Difference between the scanner volume and the one obtained by approximating the BS to a truncated cone using the subject’s manual measurements, $\Delta V_{scanner-cone,BS}$.

Figure 10

Boxplot of the resulting $\Delta L$ values, which were normalized over the subject’s height (SH). Blue boxes refer to male data and pink boxes to female data. Outliers are depicted in black with a plus symbol. (a) Difference between the BS lengths obtained from our biomechanical model and the BS lengths obtained from the literature reference, $\Delta L_{scanner-ref,BS}$. (b) Difference between the BS lengths obtained from our biomechanical model and the BS lengths obtained from manual measurements, $\Delta L_{scaner-cone,BS}$.

Figure 11

(a) Boxplot of the $\Delta D_{BS}$ computed as the difference between ${\widehat{D}}_{scan,BS}$ and ${\widehat{D}}_{ref,BS}$. The difference was normalized over the subjects’ height (SH). (b) Difference of the BSs mass resulting from our scanner-based method and the reference tables. The data were normalized over the subject’s body weight (BW).

Figure 12

(a) Histogram distribution of the normalized diameters for both the rescaled table values (red) and manual measurements (blue) for the leg BS. (b) Bland–Altman plot of the normalized diameters of the thigh BS.