Robust Method For Completely Automatic Aerial Detection Of Occluded Roads With New Initialization

A heretofore unsolved challenge is the completely automatic and accurate estimation of road boundaries in aerial images when the roads may be partially or completely locally occluded and clutter may be prevalent. In this paper we introduce a roadfinder that is effective in meeting this challenge with a new approach which improves the speed of the algorithm described by Barzohar et al.(1) In the new approach we use a new initialization for the low-level algorithm, described by Barzohar and Cooper in,(6) to obtain a completely automatic roadfinder. All the parameters that we need are estimated, such as minimum and maximum road widths, gray level variation and polarity between road and backgrounds, edge strength threshold at the road boundaries and road direction. The approach modifies the low-level algorithm, described by Barzohar and Cooper in,(6) by adding the use of a line-pair detecting procedure which identifies regions of interest (road candidates) for the low-level treatment. The roadfinder begins with a line-pair detection which identifies windows that search for road seeds, continues with one or more seeds on each long road, and then accurately estimates the remaining boundaries. The algorithm is robust to missing boundary edges on one side of the road and on both sides of the road simultaneously. (These missing edges are the result of shadows and occlusion by trees, poles, small structures, etc.) It is also robust to clutter within the road caused by cars or trucks, and to clutter resulting from two intersecting or adjacent roads. The quality performance of the automatic roadfinder in the five tested road images is based on statistical. target recognition techniques, using a manual corresponding reference for each of the five different real images. This performance is very impressive, with probability of road detection equal to 1 and probability of false alarm equal to 0, i.e. all roads in the five images were detected. The estimation of the average accuracy of our roadfinder is equal to 0.68 pixel, which expresses the average distance between the extracted roads image and the reference roads in the five images