The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
665 
Figure 2. Section of a test scene (upper image), background 
(middle image) and foreground (bottom image) 
segmented with MoG-algorithm. 
For complex outdoor scenes an adaptive approach for 
background subtraction is important, due to changing lighting 
conditions. Moreover, objects that move into the scene, but 
become static after a while, for example parking cars, should be 
classified as background. 
Three approaches for adaptive background subtraction were 
investigated: 
1) The OpenCV-implementation of a modification of the 
MoG-Algorithm: The colour of each pixel in the 
successive frames is described by multiple Gaussian 
distributions. The dominating distributions are 
interpreted as background, while weak distributions 
indicate the existence of a foreground object at the 
pixel position. 
2) The OpenCV-implementation of a Bayes decision 
framework by Li et al. (2003): The classification 
decision is based on the statistics of feature vectors. 
Stationary background objects are described by 
colour features. Furthermore moving background 
objects can be classified using colour co-occurrence 
features. 
3) Median filtering: Comparison of the actual pixel grey 
value and the Median grey value of m pixel values of 
every n-th preceding frame. If both values match, the 
existence of background is assumed. Best results with 
this approach were gained with m = 9 and n = 9. 
Figure 3. Correlation of the background estimated by three 
different approaches and the reference background. 
After an initialisation phase all tested approaches delivered 
reasonable results. Problems generally occurred if the 
background grey value and the object value are very similar. 
Figure 3 shows the correlation of the background estimated by 
the three different approaches and the “true” reference 
background. Moreover the correlation of the original image and 
the reference background is displayed. This correlation is an 
indicator for the number and size of moving objects within the 
scene. 
For the tested scenes best results were achieved with the MoG- 
Algorithm. One challenge in this approach is, after when a non 
moving object should be classified as background (e.g., person 
waiting at the traffic lights, fig. 2, middle image). Problems also 
occur in case of shadows of moving objects, since they are also 
classified as foreground objects. 
The Bayes-approach can deal with this challenge and is also 
able to deal with waving branches. In general the results of this 
approach are comparable to the MoG-results for close objects, 
but it showed weaknesses for small and distant objects. 
The Median-filtering approach is up to three times faster than 
the other approaches. The disadvantage of this approach in 
comparison to the MoG-approach, is that only one span of grey 
values is treated as background. Hence it does not reach the 
classification quality of the other approaches in case of quickly 
changing illumination conditions. 
For our further proceeding we used the MoG-Algorithm. 
Pedestrian detection 
In literature several approaches for pedestrian detection are 
described. We investigated two approaches, one based on 
background subtraction while the second approach consists of 
the application of already trained classifiers for persons. The 
later can also be applied on the original images. 
The first approach is, to group the segmented foreground pixels 
into connected components. Small components, which result 
from noise or small non-interesting objects, and which do not 
pass an area criterion, are eliminated. The remaining 
components can be divided into three categories: 
Single stand-alone persons 
Groups of persons 
Other moving objects (cars, cyclists, ...)