ICLRerNet: Lane Detection Method Based on Dynamic Lane Width and Weighted LaneIoU

Image-based lane detection plays a crucial role in autonomous driving systems, where navigation and control are heavily reliant on accurate detection of lane lines. Lane detection serves as an essential input for downstream tasks such as path planning and lane keeping within the system. In recent years, lane detection algorithms based on deep learning have achieved notable results in 2D imagery, particularly the Anchor-based 2D lane detection methods epitomized by the CLRerNet approach. Building on the CLRerNet algorithm, this paper incorporates the imaging characteristics of lane lines in 2D images, where lanes appear larger close-up and smaller at a distance, into the design of lane detection algorithms. By considering these imaging characteristics, it proposes dynamically setting lane line pixel widths and a weighted Intersection over Union (IoU) calculation method. This approach has achieved the best performance metrics on the most commonly used public datasets for lane detection, CULane and Curvelanes.