| 引用本文: | 李 政1,辜丽川2,3.驾驶区域分割与车道线检测多任务网络方法研究(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2026，43（2）：61-68 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| 目的 交通驾驶感知系统是自动驾驶技术的关键组成部分之一,该系统包含多个任务并且需要保证高实时
性和高可靠性,针对现有方法目标检测效果差,部分遮挡情况下目标区域语义分割不精确等问题,提出一种端到端
的多任务网络模型 MTNet,用于完成交通驾驶感知系统中可驾驶区域分割和车道线检测两项关键任务。 方法 在颈
部网络设计中改进特征金字塔 FPN(Feature Pyramid Network)模块以获得更高效的检测性能;在网络结构中引入
ELAN(Efficient Layer Aggregation Networks)结构和重参数化方法对原始卷积进行升级;根据不同任务特点设计了
更精准的多任务损失函数,同时提出了更有效的训练策略。 结果 通过在 BDD100k (Berkeley DeepDrive)数据集上
进行大量实验,表明可驾驶区域分割任务的平均交并比(Mean Intersection over Union,mIoU)达到 95. 6 %,车道线
检测任务的准确率(Accuracy)达到 89. 3 %。 结论 实验结果表明:MTNet 展现出对复杂背景干扰的鲁棒性,提升了
目标检测分割的精确度;无论是在白天还是夜晚,复杂还是简单场景中,都具有良好的检测分割效果。 |
| 关键词: 多任务网络 车道线检测 可驾驶区域分割 端到端训练 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Research on Multi-task Network for Drivable Area Segmentation and Lane Line Detection |
|
LI Zheng1,GU Lichuan2 3
|
|
1. School of Information and Artificial Intelligence Anhui Agricultural University Hefei 230036 China
2. Key Laboratory of Intelligent Agricultural Technology and Equipment of Anhui Province Hefei 230036 China
3. Anhui Agricultural Situation Information Perception and Intelligent Computing Engineering Research Center Hefei
230036 China
|
| Abstract: |
| Objective The traffic driving perception system is a critical component of autonomous driving technology. It
involves multiple tasks and requires high real-time performance and reliability. Existing methods suffer from issues such as
poor detection performance for small objects and imprecise semantic segmentation of target regions under partial occlusion.
To address these problems an end-to-end multi-task network model called MTNet is proposed to accomplish two critical
tasks in the traffic driving perception system drivable area segmentation and lane line detection. Methods MTNet
improved the neck network design by enhancing the feature pyramid network FPN module enabling more efficient
detection performance. The network architecture incorporated the efficient layer aggregation network ELAN structure
and reparameterization techniques to upgrade the original convolutions. Moreover precise multi-task loss functions were
designed based on the characteristics of each task and an effective training strategy was designed. Results Extensive experiments conducted on the BDD100k Berkeley DeepDrive dataset demonstrated the effectiveness of MTNet. The
drivable area segmentation task achieved an average intersection over union IoU of 95. 6 % while the lane line
detection task achieved an accuracy of 89. 3 %. Conclusion The experimental results show that MTNet exhibits
robustness against complex background interference and improves the accuracy of object detection and segmentation. The
model achieves good detection and segmentation performance in various scenarios whether during the day or at night and
in both complex and simple environments. |
| Key words: multi-task network lane line detection drivable area segmentation end-to-end training |
