A Bayesian Approach to Road Image Segmentation

It is well known that Mobile Mapping Systems (MMS) offer countless possibilities for data collection and field inventory to government agencies, public utilities, and engineering and architectural firms. Despite major progress has been made over last years in terms of sensor resolution, data rate, positioning accuracy and operational flexibility, the detection of road edges on collected the digital images is performed not yet automatically, what in some sense limits the productivity of these surveying systems. In this paper an automatic procedure for digital image segmentation will be presented, whose main goal concerns with the detection of road edges from the image sequence collected by an MMS. At the present, the application of developed method is limited to black and white digital images, acquired by a pair of mid-resolution CCD cameras (720x578 pixels), mounted on the windscreen of the vehicle (car or van). In the initial stage, each b/w image is applied the Canny edge detector and the Hough transform, in order to obtain a first approximate estimate of road edges. Then, the detection system is made more robust against any environmental condition through further image processing based on image texture analysis and the application of an extended Kalman filter. Average and standard deviation values for brightness and texture are calculated for each digital image and a Bayesian classification is performed in order to define probability maps for each component (sky, road, background) of the viewed scene. In turn, the extended Kalman filter allows to compute at each iteration the predicted positions of road edges on the next image. From these data a “virtual” image is built and then a-priori probability map is computed, by which the road segmentation can be more efficiently driven.