Vision-based vehicle guidance

The use of microprocessors to control various automobile operations is now commonplace, and further computerization can be expected as researchers extend their efforts to develop autonomous, self-guided vehicles. One of the most challenging research areas is road following, which requires the two basic functionalities of lane detection and obstacle detection. Thanks to the reduced costs of image acquisition devices and to the increasing computational power of current computer systems, computer vision has recently become a popular method for sensing the surrounding environment. The authors use an approach that extracts and localizes features of interest, thereby limiting the computation-intensive processing of images. A geometrical transform called inverse perspective mapping makes a SIMD (single instruction, multiple data) approach practical for processing data captured in stereo images. Besides contributing to obstacle detection, the left stereo image is used in lane detection. The use of a 3D surface called a horopter, moved onto the road plane by electronic vergence, makes it possible to locate obstacles and establish their distance and exact position in 3D world space. The authors describe the GOLD system, a stereo vision system developed at the University of Parma, Italy, for generic obstacle detection and lane localization. GOLD was first tested on an experimental land vehicle for more than 3,000 kilometers along extra-urban roads and freeways at speeds up to 80 kilometers per hour and is now being ported to the Argo autonomous passenger vehicle.