» 
Wo — Fh VD S T 
  
— — 
Faber, Petko 
  
adult passengers, 68 children between the ages of 1 and 11, and 18 infants. Analyzing the causes of the reported deaths it 
can be deduced that the most adult victims are unbelted and the injured childs or infants are placed in the front seat of a 
vehicle with a passenger-side airbag. And so, three different scenarios can be defined, which are of a special interest for 
an intelligent airbag deployment: 
1. the seat is occupied by a adult passenger, who is not belted (correctly) or not in a defined position on the seat, 
LD 
. the seat is occupied by a child or an infant (in a suitable child seat or rear-facing infant seat), or 
3. the seat is empty or occupied by any other object e.g. a bag or a jacket. 
In the last decade a lot of scientific groups focus their research to develop an intelligent airbag system. First solutions 
for passenger detection were already transferred into practice by different vehicle manufacturers. But, useful are systems 
which can detect and localize passengers inside a vehicle to protect the passengers. 
The literature contains a lot of aproaches to detect humans in 2-D and 3-D (e. g. (Aggarwal and Cai, 1999, Gavrila, 1999, 
Krumm and Kirk, 1998)). However, often the approaches are model-driven, bases on colour informations (e. g. (Birchfeld, 
1998, Fieguth and Terzopoulos, 1997, Sobottka and Pitas, 1996)) or the step of initialization works semi-automatically 
with restrictions on the persons position (e. g. (Schubert and Dickmanns, 1998, Yow and Cipolla, 1996)). 
In this paper we describe the research effort to develop a system to detect and localize driver and passenger inside vehicles. 
Our system reliably classifies a seat as either empty or occupied. If a seat is occupied we try to detect and localize the 
passengers’s head in 3-D. The main problem is, that we do not not know whether and where passengers are inside the 
vehicle. À sensor in the passenger seat which recognizes if the seat is occupied or not is not available. Furthermore the 
integration of a movement detection as additional feature is not possible, since a (significant) movement of a person cannot 
be assumed inevitably (e. g. sleeping passengers or children, which are limited in their freedom of movement by a child 
seat). We need to localize also such passengers. And last but not least, to make the system attractive for an integration in 
a vehicle, the costs are to be kept as low as possible, which affects also the used hardware components. Concretely, we 
can not use colour cameras. 
In section 2 we describe the used experimental setup and the developed computational structure. The step of passenger 
detection and localization is described in detail in section 3. And finally, in section 4 the feasibility of the approach is 
shown. 
2 STEREO CAMERA SYSTEM 
(80 … 100) mm 
2.1 Experimental setup > 
  
  
  
  
The used experimental setup (cf. fig. 2) consists of two b/w / x 
cameras with wide angle objectives. The sensor system is 
mounted at the windshield close to the driving mirror. Other 
positions to mount the sensor system inside the vehicle are 
possible, but are usually not accepted by the car designers. 
To take the image sequences we use only the existing light- | 
ing. Other approaches, such as infrared, radar, and ultrasonic 
are hampered by severe problems, €.g. legality, accuracy, and 
electronic smog. Also no structuring lighting is used. The 
size of the stereo images we use at the moment is 360 x 288 
[pels]. 
1200) mm 
(500 … 
  
     
  
  
driver seat passenger seat 
Figure 2: Used eperimental setup for binocular images 
of vehicles’ interior 
2.2 Computational structure 
The developed computational structure consists of the following steps: 1) the correction of distortions followed by an 
epipolar rectification of a stereo image pair, 2) a feature extraction, 3) a feature-based matching, 4) the detection and 
verification of the seat occupation, and 5) finally the detection and localization of passengers. 
The input of the developed software system are stereo image sequences and a set of calibration parameters. The output 
is defined by the information about the status of the seat occupation. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 231