Multi-class lane marking segmentation dataset for vision-based environmental perception in autonomous driving

Engineering notes

Key topics: autonomous driving, lane detection, semantic segmentation, perception, prediction. See the paper for implementation details and experimental results.

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

Advancements in lane detection, leveraging semantic knowledge, have enhanced the detection capabilities of intelligent vehicles in various traffic scenarios. However, existing lane detection algorithms encounter difficulties in reliably extracting lane instances and adapting to simultaneous variations in lane numbers. The proposed study delves into the critical realm of visual scene understanding with a focus on the semantic knowledge of lane classes. The complexity of multi-class lane marking classification is highlighted due to the eccentric, bland, and repeatable nature of lane markings, with classification relying on relative locations. In this study, an effort has been undertaken to introduce a dataset named SemSeg-Lanes, which is derived from the BDD Lane Detection dataset. This dataset encompasses 10 distinct lane marking classes specifically designed for semantic segmentation. Various baseline models, relying on established methods for semantic segmentation, are presented for this demanding dataset. Among these, real-time networks, including PPLiteSeg variants, were incorporated into the baselines and trained to achieve a mean average precision exceeding 70%. The dataset presented enables exploration of fundamental challenges in lane marking segmentation, paving the way for applications such as lane-centric activity understanding, future event prediction, and continual learning.

Links and sources

• Official / arXiv page

Comments