基于双特征提取网络的车道线识别方法
Lane line recognition method based on double feature extraction network
窦志 1孙后环 1王周利 1代远扬 1高枫1
作者信息
- 1. 南京工业大学机械与动力工程学院,南京 211800
- 折叠
摘要
为了提高复杂环境下网络的特征提取能力,提出一种双特征提取网络的车道线识别方法.首先搭建双特征提取网络,减少细节语义信息的丢失,强化模型面对复杂环境的识别能力.然后使用改进的空洞空间金字塔池化结构增大感受野,提取更为丰富的上下文信息,并结合深度可分离卷积,降低模型的计算量.最后构造通道注意力模块,重点关注有效信息较多的特征通道.经实验验证,所提方法在Tusimple数据集上准确率可达97.7％,mIoU为76.2％,单图识别时间为26.24 ms,在复杂环境下进行车道线识别时,鲁棒性较好.
Abstract
In order to improve the feature extraction ability of the network in complex environments,a lane line recognition method for double feature extraction network is proposed.Firstly,a double feature extraction network is constructed to reduce the loss of detailed semantic information and enhance the recognition ability of the model in com-plex environments.Then,the improved atrous spatial pyramid pooling structure is used to increase the receptive field and extract more rich contextual information.In addition,depthwise separable convolutions are combined to reduce the computational complexity of the model.Finally,a channel attention module is constructed to focus on feature channels with more effective information.Experimental results show that the proposed method achieves an accuracy of 97.7％and an mIoU of 76.2％on the Tusimple dataset,with a single image recognition time of 26.24 ms.When recognizing lane lines in complex environments,the proposed method demonstrates good robustness.
关键词
车道线识别/双特征提取/Swin/Transformer/通道注意力模块/空洞卷积Key words
lane line recognition/double feature extraction/Swin Transformer/channel attention module/dilated convolution引用本文复制引用
基金项目
国家自然科学基金青年基金(51804169)
出版年
2024
NETL
NSTL
万方数据