Acquisition of traffic data in an urban network is performed using a variety of sources, however, each of them has a coverage limit due to distinct reasons, such as privacy concerns or implementation costs. Since the recorded information is contained in a non-Euclidean structure and is partially unknown, we propose modelling an urban network as a road graph (i.e. with roads as nodes, and edges representing their intersections) and traffic data as incomplete graph signals to be recovered via kernel ridge regression with a kernel matrix aware of the graph topology. Evaluating distinct graph kernels based on the Laplacian matrix to recover graph signals constructed from a real-world traffic dataset (i.e., the PNeuma dataset for Athens), we show that the Laplacian RBF kernel obtains the best results in terms of smoothness and that the vehicle volume signal is well recovered with any kernel.