Multiple spatial data matching is a crucial prerequisite for data integration and interoperability, change detection and data updating. Road network matching is of great theoretical and practical significances in Navigation, Intelligent Transportation System and Location-Based Services. The paper proposes a probabilistic relaxation approach for matching urban road networks. The proposed method starts with an initial probabilistic matrix according to the geometric dissimilarities, and then integrates the effects of neighbouring roads to update the old probabilistic matrix until it is convergent to a specified small value. Finally, on the basis of the convergent probabilistic matrix, the structural similarity of each candidate pair is calculated and the corresponding rules are defined to select and refine 1: 1, 1: M and M: N matches. Two experiments of matching between Open Street Map network data and professional road network data in Wuhan and Zurich show that our method achieves a robust matching precision for large non-rigid deviation, is independent of matching direction, and successfully matches 1: 0 (Null), 1: 1 and M: N pairs.