Xu Li-Qun, David Young, David Hogg
We describe a program to construct a model of a road junction using data from a single camera. The model specifies the ground plane orientation in camera coordinates and the positions of traffic lanes, and is obtained entirely from observations of vehicle movements, with no static image analysis. At present, the model is restricted to representing straight lane segments. We describe our methods for segmentation, object tracking, ground plane estimation and lane identification. Throughout, we emphasise techniques which are computationally cheap and can be used with fairly low resolution data and a low frame rate. We nevertheless obtain a reliable model by using the statistics of large numbers of vehicle movements.
This paper is not available online