Lane Line Detection Network Based on Strong Feature Extraction from USFDNet
Zengqiang Li, Ping Lan, Qian Zhang, Liuqing Yang, Yongfeng Nie · 2024
Lane detection technology, as a critical component of autonomous driving and driver assistance systems, has received widespread attention. However, there are still many challenges in practical application. Accurately locating lane lines becomes particularly arduous due to problems such as the diversity of lanes, small occupancy, occlusion, and so on. The SFENet strategically incorporates a Path Aggregation Module before the auxiliary segmentation module. This enhancement empowers the network to delve deeper into capturing high-level semantic information within the image when performing semantic segmentation while comprehensively considering the detailed texture of the image. Furthermore, to address the issue of lane lines occupying a small portion of the image, a Multi-Scale Feature Aggregation Module is introduced before the grouping classification stage. Even on a small scale, this module ensures precise localization of lane lines and the extraction of abundant features by using dilated convolutions with varying dilation rates. Extensive experimental validations show that SFENet outperforms other algorithms in accuracy and robustness. Compared with USFDNet, SFENet achieves a 2.1 \ % enhancement in accuracy (mAP) on the CULane dataset while maintaining a processing speed of 120 frames per second.