This report presents a new stochastic geometry model for unsupervised extraction of line networks (roads, rivers, etc.) from remotely sensed images. The line network in the observed scene is modeled by a polyline process, named CAROLINE. The prior model incorporates strong geometrical and topological constraints through potentials on the polyline shape and interaction potentials. Data properties are taken into account through a data term based on statistical tests. Optimization is done via a simulated annealing scheme using a Reversible Jump Markov Chain Monte Carlo (RJMCMC) algorithm, without any specific initialization. We accelerate the convergence of the algorithm by using appropriate proposal kernels. Experimental results are provided on aerial and satellite images and compared with the results obtained with a previous model, that is a segment process called "Quality Candy".