基于行列栅格感知Transformer的车道线检测方法OA

Under low-light or uneven lighting conditions at night,road imaging suffers from low visibility of lane lines,local overexposure,and shadows.Existing lane detection algorithms primarily focus on im⁃proving detection capabilities in normal lighting environments,neglecting the degradation of road features in nighttime lighting conditions,which compromises their accuracy and robustness.To address these is⁃sues,this paper proposed a lane detection method based on a Row-Column Grid-Aware Transformer.The proposed method first employed a Light Enhancement Curve module to normalize the illumination of input images,utilizing a Generative Adversarial Network(GAN)to map low-quality images to clear ones,effectively suppressing noise and overexposure.An encoder based on ResNet34 extracted multi-scale features.The core design was a Row-Column Grid-Aware Transformer module,which explicitly modeled the spatial structure and contextual relationships of lane lines through bidirectional row and col⁃umn token encoding,enhancing the model's robustness to geometric deformations and local occlusions.The decoder consisted of a bilateral upsampling module and a confidence evaluation module,responsible for feature reconstruction and lane line existence prediction,respectively.Experimental results show that the proposed method achieves an F1-score of 76.47％in nighttime scenes on the CULane dataset,repre⁃senting an 11.09％improvement over a single-backbone network.The experimental results demonstrate that the detection accuracy of the proposed method surpasses that of current mainstream lane detection models,enabling precise and robust lane detection in complex nighttime environments.