Car, Open Your Eyes

Most traffic accidents in cities occur at intersections. It¿s therefore no surprise that researchers at DaimlerChrysler are working on supporting drivers at these high-risk locations. To do so, they are employing various methods aimed at achieving their long-term goal of creating an urban assistant.
Objekte mit der Kamera erkennen gehört in der Automation inzwischen zum Standard. Soll die Bildverarbeitung jedoch Autofahrer unterstützen, wird sie zur Königsdisziplin. In in Echtzeit muss sie aus dem Straßenbild die richtigen Objekte herausfiltern. Lesen Sie, wie das geht.
Intersections are the most critical locations with regard to urban traffic, accounting for 60 percent of all accidents in inner cities. Almost half of the accidents at intersections with traffic lights are caused by motorists driving through red lights or ignoring the right of way. Moreover, incorrect interpretations of who has the right of way account for 95 percent of all accidents at intersections with traffic signs. In view of these figures, Stefan Hahn, head of DaimlerChrysler¿s ¿Machine Perception¿ research department in Ulm, will not rest until something has been done about the problem. His team is therefore specifically concerned with city traffic and is looking into assistance systems that can quickly warn drivers of imminent danger, thus helping to prevent accidents from occurring.
Hahn points out that the focus of his work is on computer vision. The reason for this is that despite the relatively low speeds in city traffic compared to highway driving, urban situations tend to be far more complex. ¿Radar sensors can only be used to a limited degree for what we¿re trying to accomplish,¿ Hahn¿s team is therefore working with video sensors ¿ cameras ¿ that provide color or black-and-white pictures and are operated individually or in pairs as stereo systems. However, a sensor image is of no use by itself. Not until the information contained in the image is analyzed and interpreted can imminent danger be recognized. The vehicle¿s computer has very little time to accomplish this task, however ¿ real-time processing is required here. Fortunately, big advances have been made in this sector in the last few years.
20 years ago engineers had for the first time managed to develop image processing software that could recognize movement with the help of cameras. However, it took the computers an entire weekend to analyze a single second of footage. Today¿s image recognition systems, on the other hand, require only 80 milliseconds to recognize the head of a child that suddenly runs into the street from between two parked cars. This short processing time is only partially due to the enormous increase in computing power ¿ it is just as much a result of new types of image processing algorithms, which DaimlerChrysler researchers played a very important role in developing. And this is exactly the area in which Hahn¿s team is focusing its efforts.
Machine perception is a three-part process. The first step is object detection: All of the approximately ten million pixels that sensor images provide to the computer each second potentially belong to some object. If the computer and software were to try to completely process all of these pixels, it would take virtually an eternity to analyze the images. However, not all objects are equally important for traffic situations. That¿s why the system does the same thing that humans do: It perceives its surroundings in a selective manner and filters out unimportant data. This is exactly what the next two steps of the image processing procedure are designed to do.
When monitoring traffic, it is particularly important to notice movement by people or objects. This requires that the objects in question be followed over time or tracked. When tracking an object, the researchers take advantage of the fact that the position of a car or pedestrian does not suddenly change from one video image to the next. The objects have a certain ¿inertia,¿ which means the direction and speed of their movement rarely changes abruptly.
But it is not enough to simply track the object, as this does not provide any information about what kind of object it is. Here, an immediate reaction by human beings ¿ recognizing an object as a traffic light or car ¿ must be learned by the electronic image processing system with great effort. Not until it has thoroughly mastered this process can it accomplish the third step ¿ classification. To enable the computer to recognize stop signs, for example, it is shown up to 10,000 digitized images of such a sign. These are images made under various lighting conditions, by day and by night, and from all imaginable perspectives. The problem with this method is that in order to ensure that the computer knows where to find the stop sign and can also recognize it, the signs must be manually highlighted with the help of the PC mouse ¿ a very tedious business.
Anyone visiting Hahn in Ulm quickly realizes that the long-term goal of creating an urban assistant can only be achieved if a combination of new hardware and software systems are used. For Hahn, however, it will be well worth the effort: ¿After all, there¿s more at stake here than just fender-benders and property damage.¿

Dieser Artikel erschien ursprünglich im Hightech Report, Ausgabe 2/2002, und wird hier mit freundlicher Genehmigung der DaimlerChrysler AG wiedergegeben.

accomplish, to - erreichen, erfüllen
account for, to - verantwortlich sein
advance - Fortschritt
at stake - auf dem Spiel
due to - aufgrund
effort - Anstrengung
enable, to - ermöglichen
eternity - Ewigkeit
fender-bender - verbeulte Stoßstange
focus, to - konzentrieren auf
footage - Film(material)
imaginable - erdenklich
imminent - unmittelbar bevorstehend
inertia - Trägheit
intersection - Kreuzung
long-term goal - Fernziel
machine perception - Maschinelle Wahrnehmung
manner - Art
occur, to - passieren
pattern - Muster
perceive, to - wahrnehmen
property damage - Sachschaden
provide, to - zur Verfügung stellen
surroundings - Umgebung
tedious - langwierig
thoroughly - gründlich
track, to - Verfolgen
urban - städtisch

Links: http://www.engine-magazin.de