Multi-Lane Detection System Based on Segmentation Model for Autonomous Vehicles

Engineering notes

Until now, various methods and approaches have been proposed ranging from the use of traditional methods for detecting lanes to the implementation of state-of-the-art deep learning architectures. In this research, inspired by the state-of-the-art image segmentation's approaches, we have come up with novel approaches: Multi-Lane Detection System Based on Segmentation Model with applying some notable post-processing operations, which are connected component analysis and clustering methods, on the TuSimple dataset.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Nowadays, autonomous driving has recently started to attract more attention; therefore, numerous automotive companies are developing various techniques for semi-autonomous to fully-autonomous driving systems. Among these techniques, multi-lane detection is one of the critical elements of an autonomous vehicle's decision-making process. Until now, various methods and approaches have been proposed ranging from the use of traditional methods for detecting lanes to the implementation of state-of-the-art deep learning architectures. In this research, inspired by the state-of-the-art image segmentation's approaches, we have come up with novel approaches: Multi-Lane Detection System Based on Segmentation Model with applying some notable post-processing operations, which are connected component analysis and clustering methods, on the TuSimple dataset. In this study, the DeepLabv3+ architecture, which was the most effective of the several segmentation architectures utilized in the model training phase, is described. 98.98% train accuracy, 98.48% validation accuracy, 62.80% intersection over union (IoU), 89.81% precision, 85.51% recall, and 87.43% F1 score results were achieved for the model we trained. In summary, we propose an efficient multi-lane detection system that works at high performance, running at 136.7 frame per second (FPS) for the segmentation model and 16.8 FPS for overall system, can detect a variety of lanes, and can handle lane changes using a lane localization system.

Links and sources

• Official / arXiv page

Comments