Dynamic perspective transformation and feature optimization framework for robust and efficient lane detection in complex scenarios

Abstract Lane detection is a critical task in advanced driver-assistance systems (ADAS) and autonomous vehicles. Yet, existing methods face significant challenges in handling complex scenarios such as road plane variations, lane occlusions, and noise interference. To address these issues, we propose a novel approach named LaneAdapt, which combines the independently trained PerspectiveNet (P-Net) with the Rein module to achieve efficient and robust lane detection. Specifically, P-Net is pre-trained independently to learn dynamic perspective transformation matrices, overcoming the limitations of fixed transformations in non-planar road scenarios and significantly improving curve fitting accuracy and robustness. Meanwhile, the Rein module utilizes features extracted from frozen Vision Foundation Models (VFMs) and optimizes feature representations through learnable tokens and similarity mapping mechanisms, enabling precise lane instance segmentation. We also design a multi-objective loss function that integrates perspective fitting, regularization, and robustness terms to enhance the model’s stability and generalization in complex environments. Experimental results demonstrate that the LaneAdapt method achieves state-of-the-art accuracy on public datasets such as TuSimple and CULane while maintaining real-time processing performance, validating its superiority in diverse and challenging driving scenarios.