effect for this device, howeve,r some interesting symmetric 
error exists on the two spheres. 
Spehre #1 rectus - tiem. Bghero 32 sexduals - tc 
  
  
  
    
TURNS 
i 
  
  
  
(a) (b) 
Figure 10. Fitting residuals of the spheres. 
The residuals (Figure 11) show comparable results to the plane 
fitting test. It is nearly linear and about 0.1 to 0.5% of the 
measurement distance. Based on this test, we may conclude that 
Kinect has really good accuracy in from 0.7 to 2 m, and over 
that range the standard deviations are noisier and seem to be 
stochastic. The fitted point number per sphere is in range 45K 
to 1.4K 
Mean of residuals 
 Éphenedt i 
9 Sebere 2 i 
  
   
SEE Seti] 
  
  
  
Figure 11. Fitting residuals. 
In the case of feature detection and adjustment, the relative 
accuracy has a large impact (Calignano et al., 2010). The 
location variation of sphere centers describes this feature 
detection accuracy reasonably well. Figure 12 shows that the 
points’ standard deviation in the repeatability test is mostly 
under 1 cm. 
STO of center points 
  
ss 
STE Fam] 
  
Figure 12. STD of center point location variation. 
  
4. HUMAN MORPHOLOGIC MEASUREMENTS 
Typical measurement range of human morphologic is exactly 
the same as that of the Kinect. In addition, mostly only certain 
parts of the human body are examined, i.e., the face. This 
means the range up to 2 m is acceptable. Facial reconstruction 
is a fast-growing business and requires accurate human 
morphologic measurements. In addition, it is very essential to 
have a prior face model in case of plastic surgery, for example, 
after an accident. For this purpose, the Kinect gives a very good 
solution since it’s widely available and inexpensive. The 
opportunity of high speed data acquisition (30fps) is also a 
benefit of this device, as it helps avoid errors on fast changing 
(mimics) and moving human body. The accuracy is the only 
limitation of this device, though it can be increased by camera 
calibration (including special scaling factor) and using of 
multiple devices. The post processing and model generation 
should be done in a special way (Figure 13); some points can be 
dropped and key points should be used with higher weight 
(Varga et al., 2008). 
  
Figure 13. Key points (red) locations on a human face (Varga 
et al., 2008). 
5. CONCLUSION 
In our experiences, the Kinect sensor has shown good and 
consistent performance. The tests confirmed that rather good 
quality 3D imagery can be acquired in close range by this 
absolutely inexpensive sensor. The availability of several open 
source tools and the existence of an active user community 
make the integration of the Kinect sensor fairly simple, 
including basic data processing tasks too. While the Kinect is 
not a typical mapping sensor, its performance level makes it 
feasible to several applications, like human morphologic 
measurements. 
6. ACKNOWLEDGEMENTS 
The research work is funded by Hungarian Scientific Research 
Fund (OTKA no. 73251) 
   
   
   
   
  
   
    
   
   
    
    
    
    
   
    
  
   
   
   
    
   
   
   
    
  
   
   
  
  
  
  
  
  
  
    
   
    
  
REFER 
Aoki, Y 
Simulati 
physic-b 
Photogr 
Supplem 
Calignar 
facial r 
features 
ISPRS, 
XXXVII 
Github, | 
Kahlmar 
for inci 
SwissRa 
Engineer 
136-141. 
Khoshell 
Data. IS] 
Vol. XX 
Konolige 
Kinect 
kinect cz 
Microsof 
(10 Apr.