Faber, Petko 
  
Stereo image + 
calibration parameter 
  
Distortion correction + 
epipolar rectification 
Ÿ 
( Feature extraction 
  
  
$ 
Feature-based matching 
1 
Seat occupation detection 
  
    
  
  
Passenger detection and 
Status of seat occupation 
Figure 3: Computational structure. 
Distortion correction and epipolar rectification: The input of this module is defined by stereo images and the calibra- 
tion parameters. The calibration of the stereo system is a priori done by a test field based camera calibation performed 
off-line. The output is an epipolar stereo image, where corresponding points have the same vertical coordinates in 
both images, thus lie on the same scan-line. The objective of this module is to map the stereo images into a virtual 
stereo image, which have aligned viewing directions, identical image planes and parallel image coordinate systems. 
Feature extraction: The input ofthe feature extraction is an epipolar stereo image without any distortions. The output 
of this module is a list of image points extracted in both stereo images. The determination of significant features 
is the basis for all following steps. Here, we use significant edge points as features, whereby a significant point 
is defined by the following useful requirements: the points should be sufficiently different to their neighborhood, 
invariant to geometrical and radiometrical transformations, seldom, interpretable, and the extraction of points should 
be insensitive to noise. The implemented algorithm is a one-dimensional version of the one published by (Fórstner 
and Gülch, 1987). 
Feature-based matching: The objective of this module is to find corresponding feature points in order to reconstruct 
their 3-D position. The input of the matching is the list of points of interest in both images. The output is a list 
of corresponding points with their similarity as additionally attribute. The matching is realized in two stages. Based 
on a list of candidate matches, i. e., putative corresponding points, using a similarity measure, determined in the first 
step, the correspondences are made unique by concictency checks in the second step. The resulting list obeys the 
principles of exclusiveness and similarity. Finally, the 3-D coordinates of the matched image points are determined 
by intersecting the image rays passing through the optical center and the image points. 
Seat occupation detection: The input consists of the list of corresponding points. The output of this module is the 
decision whether the seats are empty or occupied. If a seat is classified as empty one obtains the form and position 
of the seat. In the case a seat is occupied we do not distinguish between passengers and any other objects. The 
seat occupation detection is the first module in our computational structure, which supplies usable information to 
an intelligent airbag deployment. The knowledge of position and form of driver and passenger seat facilitates the 
classification of seat occupation substantially. The selected seat model, an elliptical cylinder, can be easly replaced 
by more sophisticated one in order to improve the results if necessary. The developed procedure is described in 
details in (Faber, 2000). 
Passenger detection and localization: The input is the list of corresponding points and the available information about 
the status of seat occupation. The output of the last module at the moment is defined by the information about a 
possible seat occupation by a human. Thereby, if a seat is occupied by a human, the information about position and 
orientation of the human's head as the most distinguishing feature of the body are available. The head is approximated 
by an ellipsoid, specified by geometrically interpretable parameters. 
A detailed description of the developed procedure is given in the following section. 
3 PASSENGER DETECTION AND LOCALIZATION 
The detection and localization of passengers inside a vehicle is the most important step within an intelligent airbag 
deployment. On the basis of the determined position of a passenger inside the vehicle it should be possible to control the 
airbag deployment in the case of an accident directly. 
3.1 Data 
The shape of the human's head can be well approximated by a rigid, three-dimensional model because the head is nearly 
rigid and roughly symmetrical. The head is modeled by an ellipsoid as mentioned with the geometrical restrictions of a 
  
N 
o 
t3 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 
  
SL