电子学报 ›› 2021, Vol. 49 ›› Issue (7): 1379-1385.DOI: 10.12263/DZXB.20201375

• 学术论文 • 上一篇    下一篇

基于事件驱动的车道线识别算法研究

徐频捷1,2, 陈逸杰3, 李之南4, 赵地1   

  1. 1.中国科学院计算技术研究所, 北京 100190
    2.中国科学院大学, 北京 100049
    3.北京邮电大学信息与通信工程学院, 北京 100876
    4.北京邮电大学国际学院, 北京 100876
  • 收稿日期:2020-12-01 修回日期:2021-02-22 出版日期:2021-07-25 发布日期:2021-08-11
  • 作者简介:徐频捷 男,1997年生于江西上饶.现为中科院计算技术研究所硕士研究生.E‑mail:xupinjie19g@ict.ac.cn
    陈逸杰 男,1999年生于云南昆明.现为北京邮电大学信息与通信工程学院本科生.E‑mail:yj‑chen@bupt.edu.cn
    李之南 男,1998年生于新疆乌鲁木齐.现为北京邮电大学本科生.主要研究方向为深度学习加速器,机器学习.E‑mail:a710859414@bupt.edu.cn
    赵 地(通信作者) 男, 1978年生于湖南岳阳. 现为中科院计算技术研究所副研究员,主要研究方向为类脑计算和深度学习.E‑mail: zhaodi@ncic.ac.cn
  • 基金资助:
    国家自然科学基金(61420106013);北京市自然科学基金(4161004);北京市科技项目(Z161100000216143)

Research on Event‑Driven Lane Recognition Algorithms

Pin-jie XU1,2, Yi-jie CHEN3, Zhi-nan LI4, Di ZHAO1   

  1. 1.Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China
    2.University of Chinese Academy of Sciences,Beijing 100049,China
    3.School of Information and Communication Engineering,Beijing University of Posts and Telecommunications,Beijing 100876,China
    4.International School,Beijing University of Posts and Telecommunications,Beijing 100876,China
  • Received:2020-12-01 Revised:2021-02-22 Online:2021-07-25 Published:2021-08-11

摘要:

动态视觉传感器(Dynamic Vision Sensor,DVS)相比于传统彩色相机有更高的时间分辨率、动态范围,且功耗更低、带宽更低,在自动驾驶领域有很好的应用前景,因此吸引了越来越多研究者的注意.然而由于事件驱动的数据是异步的且缺少一种统一的表示形式,在复杂的交通场景下,以车道检测为代表的基于事件驱动数据的交通场景分割任务难以应用传统的语义分割算法.针对以上问题,本文提出了一种三通道的事件数据编码方式,综合考虑事件数据的时空特征,将其作为卷积神经网络的输入;提出了一种基于编解码模型的事件数据车道检测算法,在基于事件驱动的车道线检测数据集DET上,本文方法的mIoU(mean Intersection over Union)达到了58.76%,比基准方法提高了4.4%.

关键词: 事件驱动, 卷积神经网络, 车道线检测, 编解码模型, 语义分割, 动态视觉传感器, 事件表示

Abstract:

Compared with the traditional color cameras, the dynamic vision sensor, a type of event?based sensor, has higher time resolution, dynamic range, lower power consumption and lower bandwidth requirements. It has good application prospects in the field of automatic driving, which attracts more and more researchers’ attention. However, event?driven data is asynchronous and lacks a unified representation. At the same time, in the complex traffic scenario, the traditional semantic segmentation model is difficult to be applied to the event?driven data?based traffic scene segmentation task, for instance, the lane detection task. In view of the above problems, our study proposes a three?channel encoding method for event data, which is successfully used as the input of convolution neural network by considering the spatio?temporal characteristics of event data comprehensively. This paper also proposes a lane segmentation algorithm based on encoding?decoding model, which is superior to the traditional event?based lane line segmentation. On the DET data set, with mIoU(mean Intersection over Union) as the evaluation index, this paper reaches 58.76%, which is 4.4% higher than the benchmark.

Key words: event?based, convolution neural network, lane detection, encoder?decoder model, semantic segmentation, dynamic vision sensor, event representation

中图分类号: