Lane line detection is a key technology to realize autonomous driving, which is a fundamental and challenging task in autonomous driving. In this paper, a semantic segmentation algorithm for lane lines based on multi-scale deep feature fusion is proposed. By analyzing the spatial structural properties of continuous elongated lane lines, we design a multimorphic CASPP module, which combines the mutual quality null rate with 1D convolutional branching to enhance the context-awareness of elongated linear features. The DeepLab-ERFC model is further constructed to introduce the enhanced boundary learning of ER Loss based on Hausdorff distance, combined with dynamic gradient correction to alleviate the category imbalance problem, and optimize the prediction boundary using the post-processing of fully-connected CRFs. Experiments on TuSimple, VPG and tvtLANE datasets show that the model significantly outperforms mainstream methods in both accuracy and speed, with average intersection and merger ratios of mIoU reaching 64.62%, 68.79% and 64.62%, respectively, which is an improvement of 2.12-8.31 percentage points over models such as DANet and PSPNet. In terms of real-time, the inference speed reaches 89.34 FPS, which is more than 2.6 times higher than the comparison model. The ablation experiment verifies the effectiveness of the multi-module synergistic optimization, with the CASPP module increasing the mIoU by 5.75%, the ER Loss with gradient correction by a further 6.86%, and the CRFs postprocessing finally pushing the mIoU to 64.62%. Under extreme scenarios (e.g., sudden changes in tunnel light, vehicle occlusion, rain and snow interference), the average accuracy of the model improves by 3.8-21.3 percentage points over the suboptimal method, demonstrating strong robustness. The model constructed in the article significantly improves the accuracy, stability and real-time performance of lane line detection, thus realizing safer and more efficient autonomous driving technology.
