With increasing traffic density the demand for higher security and greater comfort for the car driver arises. Therefore new technologies are required. One of these is computer vision, which can be used to support the driver in complex situations in order to increase his security and comfort. A possible function to be used in future driver support systems is automatic object and lane marker detection. For these applications active sensors like radar or lidar show significant problems in the near vicinity of the sensor. As an alternative video can be applied for object detection, which is the subject of this paper, for longitudinal control. The problem of detecting objects outside the road plane can be solved by using the 3D information from two images of the same scene recorded from different perspectives by a stereo rig. Here the surface of the road is modelled by a plane. Then a set of eight collineation parameters can be determined which maps any point of the road from one perspective of a stereo rig to the other perspective. Applying this mapping for every pixel of one image, this image appears to be recorded from the other perspective. Obviously this is only valid for points which belong to the plane of the road. For points outside the road plane this mapping fails, resulting in a high energy in the difference image from which an object mask can be derived. From this object mask individual objects can be extracted and tracked in time.
展开▼
