In this study, the traffic cone used to construct the temporary road is taken as the research objective, and the three-dimensional point cloud data of the temporary road collected by multi-line LiDAR is taken as the input. A graph neural network model based on graph theory is proposed, which can realize the segmentation of point cloud data and improve the learning effect of the model on the disordered point cloud data. Take the driverless formula car as the experimental platform, train and test for traffic cone, the experimental results show that the segmentation accuracy of the graph neural network model reaches 88.6%, which is about 10% higher than that of the PointNet model. In addition, the model also has a certain generalization ability under sparse LiDAR point cloud data, and has good applicability.