Curb detection and mapping are of great importance to ensure the safety and efficiency of intelligent vehicles. However, it remains challenging because shape estimation under noise and outliers is not well addressed in real traffic scenarios. In this paper, an efficient curb detection and mapping algorithm is proposed to achieve an accurate representation of curb shape. More specifically, an iterative Gaussian process regression (iGPR) is introduced, where each candidate point is verified multiple times. Then iGPR is employed in shape estimation of both road profile and curb, which serves as the backbone unit in curb candidate detection. During this process, the input 3D point cloud is segmented into road and obstacles, and potential curb points are selected by evaluating physically interpretable curb features. Finally, the proposed iGPR is validated and tested on two large-scale, complex urban datasets under real traffic scenarios. Experimental results show that the proposed iGPR achieves better performance than several state-of-the-art algorithms.