Rectangular Road Marking Detection with Marked Point Processes

We propose in this article an energy minimization based approach to detect dashed lines of road markings from very high resolution aerial images (< 20 cm). The strategy we presented in a previous work was based on the grouping of image features (i.e. segments) extracted from the images (reference, n.d.), but suffers from a few problems such as occlusions mainly due to cars or changing in illumination conditions like shadows areas produced by trees or buildings or low contrast due to worn out markings. In order to obtain a more robust and reliable extraction of those objects, we have to find another approach which can take into account those observations. In this context, we use marked point process which are random variables whose realization are configuration of geometrical shapes (rectangle in our case). We use an energy formulation based on both an external and internal terms. A RJMCMC sampler coupled with a simulated annealing is used to find the optimal object configuration according to the proposed energy. Some results are shown on various aerial images in urban areas.