Stereo for 2D visual navigation

We propose a new stereo method for 2D navigation without depth search and metric camera calibration. Although there is such a stereo method using the constraint that a vehicle moves on flat planes, the conventional method has the following drawbacks. First, it lacks of an effective contrivance to deal with camera vibrations and road inclination. Second, there is no measure to assess the danger of detected obstacles. To solve these problems, we develop a domain-specific stereo method which utilizes attributes of roads. We introduce a new camera model that is valid in road scenes in order to simplify the epipolar geometry giving a useful relationship between a pair of images. Furthermore, extraction of two parallel lines marking the driving lane enables us to correctly detect obstacles and even to estimate the degree of the danger. The proposed method is implemented on a Pentium II 450 MHz PC and the current operating frequency is 15 Hz for 320×240 images without MMX instructions. Through experiments we demonstrate that our method is applicable to a variety of road scenes