Skip to Main Content
A new stereovision-based method for the road lane detection and 3D geometry estimation is presented in this paper. The proposed approach is based on a recognition algorithm driven by a statistical model of the 3D road lane, projected in both stereoscopic images. First, the model is initialized thanks to a training stage. The model is then updated iteratively, from successively extracted image features. After each iteration, the detection of the next features, in any of the two images of the stereoscopic pair, is driven by the features already detected. The parameters of the road lane, such as width, horizontal and vertical curvature, roll, pitch, and yaw angles, are estimated. The variance of each parameter is also estimated, and is minimized through the estimation process. Unlike previous proposed approaches, no disparity map is required : the matching of the image features is directly obtained as a result of the model update. Thus, computing time is low. Experiments from computer-generated and real images are carried out to assess the efficiency and accuracy of the method.