RGB sensor (2D 
    
Laser emitter E 
Laser sensor (3D camera) 
Figure 1. Kinect XBOX 3600 sensor, including 2D and 3D 
imaging sensors. 
Microsoft provides an SDK (Windows 7, Visual Studio 2010 
Enterprise, and DirectX) to support application developments 
(Microsoft). Kinect has a default measuring range of 0.3 m and 
to 3.9m (no ambiguity), which can be extended; our 
experiences indicate that up to 10 m range, reliable depth 
images can be acquired. The available open source drivers 
provide additional opportunity to acquire raw data and a very 
powerful SDK is also available. In our investigation the 
SensorKinect driver (Github) was used with OpenNI (OpenNI), 
and all the subsequent processing was done in Matlab. A typical 
2D and 3D image pair is shown in Figure 2. 
  
(a) (b) 
Figure 2. Kinect 2D (a) and 3D (b) images. 
3. ACCURACY TEST 
Kinect is an inexpensive, low-end, commercial device, yet, it 
has the potential for mapping applications, including human 
morphological measurements. Based on statistical analysis, the 
error budget of the sensors should be determined by various 
tests. 
3.1 Sensor Repeatability Test 
The repeatability of the range measurement is an essential 
aspect of using depth imaging sensors, as it provides the 
assessment of the ranging precision in short term. To determine 
the sensor repeatability performance, a planar target was 
imaged from a distance ranging from 0.5 m to 5 m in 0.1 m 
steps. Figure 3 shows 3D (depth) images of the target with and 
without the environment. 
The measurement was repeated six times, so a total of 46 x 6 
images were acquired and processed. The planar target has a 
size of 180 cm x 60 cm, so its FOV in the image changes a lot. 
Consequently, the number of points obtained by the 3D sensor 
from the reference planar target varies over a large range, from 
200K down to 10K. 
Beaicsdasy. 30cwi st Fsegihi: 3oGove ei 
  
(a) (b) 
Figure 3. Pseudo color 3D images taken at 270 cm ranges; 
entire image (a) and image of the planar target extracted (b). 
In the first step, the standard deviation was computed on a point 
basis for each distance. The repeatability results, shown in 
Figure 4, clearly indicate a near linear dependency on the range. 
The overall performance for the whole range is lower than 
0.596, which is quite excellent compared to earlier Flash LiDAR 
results (Kahlmann ef a/., 2006). 
Repeatability on point basis. 
  
— ; ; 
5 i ; 
£m Wan TG ANG TG 2006 3605 4000 4500 5060 
imm] 
Figure 4. Repeatability result. 
Figure 5 shows 2D error surfaces at two ranges, 1 m and 2.3 m, 
respectively. While the overall residual error numbers are small, 
their spatial distribution is somewhat unusual. Note that the 
circular pattern is caused by the distance calculation method, as 
described in (Khoshelham, 201 1). 
    
(a) (b) 
Figure 5. Distribution of fitting errors at ranges (a) 1 m and (b) 
2.31n. 
Based on the six measurement sets, the fitting plane residual 
errors were calculated and basic statistical parameters were 
determined, including maxima and STD for each range. Figure 
5 shows the results, including a maximum error envelope and 
the STD (6a) as well as normalized values (6b). The results 
clearly indicate good accuracy performance, as at the shortest 
  
  
    
  
object di 
error is 1 
mm and 
should t 
method, 
range, th 
error Is à 
test are e 
Figu 
3.2 Sph 
Based on 
was perf 
radius of 
4 m with 
10 times 
accuracy 
collected 
directly c