Collaborativeperception is crucial for addressing challenges related to occlusions and sensor limitations in autonomous driving. However, it necessitates accurate relative pose transformations, which are challenging to obtain with existing navigation systems. Therefore, we propose a distributed object-level collaborative perception framework called GraphPS. The graph matching method is designed to identify correspondences between co-visible bounding boxes from different perspectives, allowing for the calculation of optimal pose transformations to correct localization errors. Furthermore, to address issues of insufficient extension of the perception range and unreliable collaboration due to the lack of co-visible objects in multi-agent object-level fusion, we introduce a framework for multi-hop perception fusion based on graph pair sequences derived from graph similarity computation. This framework extends the perception range while ensuring collaboration reliability. We conduct extensive experiments on simulated datasets in V2X and V2V scenarios. Our method demonstrates robust performance under varying degrees of position and heading errors, surpassing state-of-the-art baselines in accuracy. GraphPS substantially enhances the object association accuracy and relative pose calibration precision of the baseline methods in multi-agent collaboration, showcasing its practicality in real-world complex traffic scenarios.