Real-time Vision-based Intersection Detection For A Driver's Warning Assistant

Abstract

In urban traffic situations image sequence evaluation can provide information about the environment of a car to a driver's warning assistant and thus enables it to give warning messages if the driver approaches an intersection too quickly. The image sequence evaluation procedure has to detect an intersection in realtime in a sufficiently large distance (i.e. about 30 m) with a camera mounted inside a passenger car at 1.30 m of height. It additionally detects the road borders in order to estimate the pitch angle and to establish search areas for the intersection detection. The intersection is detected with a model-based approach, in which the model contains information about usual intersection markings such as dashed or solid lines. We start with an initial model of the road and then derive first search windows for the borders. Inside these windows, edges with appropriate angles are extracted, which are then used to build the road borders. The search window for the intersection is deter mined by the estimated borders and two horizontals narrowing down the scope to distances of interest. Horizontal structures are extracted, labeled as candidates for upper and lower edges of road markings and assembled so that they match the intersection model as exactly as possible. The extent to which the hypothesis resembles the model determines the value of a scoring function. The best hypothesis is considered as the detected intersection, provided it's scoring function exceeds a threshold