基于LSTR和Vit-CoMer骨干的车道线检测方法
Lane detection method based on LSTR and ViT-CoMer
郑剑 1郭亦畅1
作者信息
- 1. 江西理工大学信息工程学院 赣州 341000;江西理工大学宜春新能源锂电研究院 宜春 336023
- 折叠
摘要
针对LSTR算法在实际应用中存在的提取特征尺度单一及缺乏对车道局部特征有效捕捉的问题.本文首次将Vit-CoMer骨干网络用于车道线检测任务中,提出LSCoMer车道线检测模型.首先,在特征提取网络后使用MRFP丰富多尺度特征,提高检测精度;其次,在Transformer结构的开始和结束位置集成CTI模块,以促进CNN的局部特征与Transformer的全局特征之间有效融合,强化后者在局部细节上的敏感性.实验结果表明,本文方法在TuSimple数据集上准确率为96.68%,较原LSTR方法提升0.5%且显著优于PolyLaneNet等同类方法,在CULane数据集中,本文方法F1 分数比LSTR方法提升3.02%.
Abstract
Addressing the limitations of the LSTR algorithm in practical applications,particularly its single-scale feature extraction and lack of effective capture of local lane features,this paper introduces the Vit-CoMer backbone network for the first time in lane detection tasks,proposing the LSCoMer lane detection model.Initially,the model employs a MRFP module after the feature extraction network to enrich multi-scale features,thereby enhancing detection accuracy.Additionally,a CTI module is integrated at both the beginning and the end of the Transformer structure to promote effective fusion between CNN's local features and Transformer's global features,enhancing the latter's sensitivity to local details.Experimental results indicate that this method achieves an accuracy of 96.68%on the TuSimple dataset,which is a 0.5%improvement over the original LSTR method and significantly outperforms similar methods like PolyLaneNet.On the CULane dataset,our method improves the F1 score by 3.02%compared to the LSTR method.
关键词
车道线检测/LSTR算法/Transformer/多尺度特征Key words
lane detection/LSTR algorithm/Transformer/multi-scale features引用本文复制引用
出版年
2024
NETL
NSTL
万方数据