How to find a path in the absence of traffic markings? A multi-scenario approach to virtual lane construction

Lane marking recognition technology is a critical component of the environmental perception system in autonomous vehicles. It is the cornerstone for achieving safe and compliant driving. It not only supports lane-keeping and path-following functions in autonomous vehicles, but also serves as the core and key for enabling autonomous lane changing and navigation assistance. Currently, the application of lane marking recognition technology in complex traffic scenarios still faces many challenges. For instance, in older urban areas, the lane markings may be unclear, and on congested roads, at night, or on curves, lane markings may be obstructed. When confronted with these scenarios where traffic markings are missing, existing algorithms often perform sub optimally. To address the above issues, this paper proposes an improved network method based on the ResNet-18 convolutional neural network for virtual lane construction. The lane reconstruction process is transformed into an anchor point and cell-based segmentation and selection problem, which facilitates the acquisition of global information, thus solving the problem of path-following in autonomous driving when traffic markings are missing in complex scenarios. Finally, model validation was conducted on four lane marking datasets: Tusimple, CULane, CurveLanes and a self-collected dataset. The validation results show that the proposed method achieves high accuracy and fast processing speed, yielding excellent performance on both datasets. Furthermore, compared with other methods, our approach demonstrates comprehensive advantages.