Local registration and deformation of a road cartographic database on a SPOT satellite image

Abstract

Herein, we propose a new method to locally register cartographic road networks on SPOT satellite images. This approach is based on Markov random fields (MRF) on graphs. Since the cartographic and image data are obtained from different sources, the noises degrading these data are of different nature. Cartographers also introduce, in the generalization process, distortions in the road map in order to emphasize some details of the road. This can create important differences between the map data and the ground truth. The proposed algorithm aims at correcting the error due to noise and generalization, hence increasing the accuracy of the road map. The first step of the method is to translate the road network into a graph where the nodes are characteristic points of the roads (e.g., crossroads). The random variable or descriptors are defined by the nodes position. The edges are defined by the roads joining these points. Then, local registration is performed by defining a model in a Bayesian framework. The solution is obtained by computing the maximum a posteriori (MAP). The posterior probability is assumed to be a product of two probabilities, the prior of the network and the likelihood of the map, each depending on the image data. Both are Markov Random Field probabilities. The likelihood of the registered map is the probability of a network configuration given the map data. It is a measure of a global resemblance between the two. We use geometrical measures, euclidean distances and angles, to build this probability. The prior consists of two terms, both depending on the image data. The dependance exists through the fact that between two connected nodes, we compute a best path, thanks to a dynamic programing algorithm, minimizing a cost function based on image gray levels, curvature and gradient information. The first term of the prior penalizes configurations for which different roads overlap each other, and the second term depends on gray level statistics along these paths. We run a simulated annealing algorithm to optimize the proposed model. The tests are done on one real image data extracted from SPOT satellite images, and artificially noisy cartographic data (translated, rotated or randomly deformed network). We present some results showing a good global registration, but also accurate correction of local distortions