A pseudo-hierarchical graph matching procedure dedicated to object recognition is presented in this paper. From a single model image, a graph is built by extracting invariant local features and linking them according to a so-called proximity rule. The resulting graph presents several interesting properties including invariance to scale, robustness to various distortions and empirical linearity of the number of edges with respect to the number of nodes. The matching process is made hierarchical in order to increase both speed and detection performances. It relies on progressively incorporating the smaller model features as the hierarchy level increases. As a result, even a matching between graphs containing thousands of nodes is very fast (a few milliseconds). Experiments demonstrates that the method outperforms state-of-the-art specific object detectors in terms of precision-recall measures and detection time.