Inverse perspective mapping （IPM） of pavement images is a prerequisite for image-based vehicle distance perception and pavement damage measurement. The traditional static IPM methods have the problem that the transformation parameters cannot be dynamically adjusted， and the existing dynamic IMP methods are highly dependent on the information such as road lane lines and textures，which often lead to suboptimal correction of perspective distortion. To solve these problems， this study proposed a dynamic IPM method for pavement images based on depth-camera semantic segmentation and 3D plane fitting. First， a semantic-segmentation model was used to extract pavement regions from RGB images， and 3D plane fitting was performed on the corresponding point-cloud data within the pavement regions， eliminating the interference of non-pavement point clouds on pavement fitting. On this basis， using pavement information and the spatial positional relationship between the camera and the pavement， the relative pose of the camera with respect to the pavement was calculated through a camera-pose-estimation method. Finally， based on the imaging relationship of the pavement under different camera poses， a constructed pavement-image IPM model was used to correct perspective distortion from the original image to any reference point. Simulation experiments show that the perspective distortion correction error of the proposed method is stable at 10^-2 mm when the camera pose has common variations， which is better than the current advanced IPM methods， demonstrating that the proposed method effectively improves the quality of pavement image IPM. Real-world experiments further validate the effectiveness of the proposed method.