Enhancing Inverse Perspective Mapping for Automatic Vectorized Road Map Generation

TL;DR

The paper tackles the challenge of cost-effective, high-precision vectorized road map generation using monocular cameras. It enhances inverse perspective mapping (IPM) by jointly optimizing the IPM matrix, vehicle poses, and map elements, while representing lanes with Catmull‑Rom splines and other markings as bounding-box polygons. Key contributions include uncertainty-guided selection of lane-control points (IPM error guidance), z-axis coordinate optimization, and a PPSR-based lane optimization that improves robustness to erroneous lane points, all validated in port and public-road scenarios with near-centimeter accuracy and IPM calibration parity with manual methods. The approach reduces data volume through spline-based lane representations and demonstrates practical impact for scalable, low-cost HD map generation in real-world autonomous systems.

Abstract

In this study, we present a low-cost and unified framework for vectorized road mapping leveraging enhanced inverse perspective mapping (IPM). In this framework, Catmull-Rom splines are utilized to characterize lane lines, and all the other ground markings are depicted using polygons uniformly. The results from instance segmentation serve as references to refine the three-dimensional position of spline control points and polygon corner points. In conjunction with this process, the homography matrix of IPM and vehicle poses are optimized simultaneously. Our proposed framework significantly reduces the mapping errors associated with IPM. It also improves the accuracy of the initial IPM homography matrix and the predicted vehicle poses. Furthermore, it addresses the limitations imposed by the coplanarity assumption in IPM. These enhancements enable IPM to be effectively applied to vectorized road mapping, which serves a cost-effective solution with enhanced accuracy. In addition, our framework generalizes road map elements to include all common ground markings and lane lines. The proposed framework is evaluated in two different practical scenarios, and the test results show that our method can automatically generate high-precision maps with near-centimeter-level accuracy. Importantly, the optimized IPM matrix achieves an accuracy comparable to that of manual calibration, while the accuracy of vehicle poses is also significantly improved.

Figures (11)

• Figure 1: An example of the generated vectorized HD map. At the turn, there is a pedestrian crossing at both the front and rear of the UGV, which are marked by blue polygons. The corner points of them are stored in the map and marked in cyan in the point cloud. The lane lines are generated as Catmull-Rom splines. The control points of them are shown as red spheres in the point cloud, the sampled lane line points are marked as black points.
• Figure 2: The workflow of our proposed system for automatic vectorized road mapping. The system can be divided into two parts. The front end is responsible for generating temporary maps and associating newly detected markings with those that have already existed on the map. When the data association succeeds, the back end jointly optimizes the map, camera extrinsics, and vehicle poses. The optimized map data updates the database, and the optimized camera extrinsic is fed forward to the front end.
• Figure 3: Different types of ground markings that possess various corner types. From left to right are zebra crossings, straight arrows, left turn arrows, speed limit signs, and lane lines. Each rectangle included in the zebra crossing contains four corners, which are not all marked in the figure. The speed limit signs do not include corners at all.
• Figure 4: Examples of ground markings segmentation. The type of markings can be self-defined in the system.
• Figure 5: This is a visualization diagram of our general optimization objectives. The red markers in the figure represent the positions of map elements projected onto the camera plane from the generated map, while the green markers represent the results of instance segmentation. Our optimization goal is to reduce the offset between them (blue arrows).