Data-driven extraction of curved intersection lanemarks from road traffic image sequences

Abstract

Segmentation of optical flow fields, estimated by spatio-temporally adaptive methods, is - under favourable conditions - reliable enough to track moving vehicles at intersections without using vehicle or road models. Already a single image plane trajectory per lane obtained in this manner offers valuable information about where lane markers should be searched for. Fitting a hyperbola to an image plane trajectory of a vehicle which crosses an intersection thus provides concise geometric hints. These allow to separate images of direction indicators and of stop marks painted onto the road surface from side marks delimiting a lane. Such a ‘lane spine hyperbola’, moreover, facilitates to link side marks even across significant gaps in cluttered areas of a complex intersection. Data-driven extraction of trajectory information thus facilitates to link local spatial descriptions practically across the entire field of view in order to create global spatial descriptions. These results are important since they allow to extract required information from image sequences of traffic scenes without the necessity to obtain a map of the road structure and to make this information (interactively) available to a machine-vision-based traffic surveillance system. The approach is illustrated for different lanes with markings which are only a few pixels wide and thus difficult to detect reliably without the search area restriction provided by a lane spine hyperbola. So far, the authors did not find comparable results in the literature.