欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 计算机视觉 • 上一篇    下一篇

基于车载视觉的驾驶员后视镜查看行为检测

  

  1. 1. 厦门理工学院机械与汽车工程学院，福建 厦门 361024；
    2. 厦门理工学院福建省客车先进设计与制造重点实验室，福建 厦门 361024；
    3. 北京理工大学机械与车辆学院，北京 100083
  • 出版日期:2018-06-30 发布日期:2018-07-10
  • 基金资助:
    国家自然科学基金项目(61401382，61104225)；福建省自然科学基金项目(2015J01672)

Apply Vehicle Vision to Detect Driver’s Rearview Mirror Watching Behaviors

  1. 1. School of Mechanical and Automotive Engineering, Xiamen University of Technology, Xiamen Fujian 361024, China;
    2. Fujian Provincial Key Laboratory of Bus Advanced Design and Manufacture, Xiamen University of Technology, Xiamen Fujian 361024, China;
    3. School of Mechanical and Vehicular Engineering, Beijing Institute of Technology, Beijing 100083, China
  • Online:2018-06-30 Published:2018-07-10

PDF (PC)

289

摘要： 车辆转向时驾驶员的后视镜查看行为是行车安全的必要措施之一，但目前关于该
行为的检测技术应用尚属空白。为督促驾驶员在车辆转向时及时查看后视镜留意车辆侧后方的
交通情况，基于车载单目视觉与图像处理技术提出一种自适应检测方法。首先，设计帧差搜索
分割算法自动实现车辆启动期间的驾驶员脸颈初始区域定位和灰度初值计算，摆脱算法对驾驶
员信息的依赖；设计胀缩分割算法快速实现车辆行驶期间的驾驶员脸颈区域定位和灰度均值计
算。其次，提取脸颈外轮廓并定义了一种由颈部轮廓基点垂线划分的左右面积比特征参数，分
析表明其受驾驶员头部姿态显著影响。最后，结合驾驶过程的眼动凝视数据揭示了该特征参数
在驾驶员查看后视镜过程中的累积概率局部峰值分布规律，并提出一种基准特征值实时估算方
法和后视镜查看行为阈值判定原理。实验结果表明，该方法适应于不同脸型，具有良好的抗干
扰能力，综合检测准确率达96.1%。

关键词: 交通安全, 车载视觉, 驾驶员, 脸颈外轮廓, 后视镜查看行为

Abstract: The driver’s rearview mirror watching behavior is one of the necessary steps for driving
safety when the vehicle is turning, however, the detection technology or application of this behavior
is still absent. Thus an adaptive detection method of the drivers’ rearview mirror watching behaviors
during the vehicle steering process was presented in this paper with the help of vehicle vision and
image process technology for safety monitoring and reminding. A frame spatial gradient differences
searching algorithm was designed to complete the initial parameters’ learning work on both the
drivers’ face and neck regions when the vehicle engine was fired, while a expand-contract searching
algorithm was invented to accomplish a fast recognition when the vehicle was moving. Contours of
the driver's face and neck parts were extracted without segmentation. An area ration between left and
right parts of the contours separated by a vertical line passing through the base point of neck contour
was defined as a characteristic parameter. By analyzing the drivers’ eye movement data during driving, a discipline called local peak value distributing of the parameter’s cumulative probability was
uncovered, which helped to build a real time eigenvalue reference estimation method and a threshold
judging principle of the drivers’ rearview mirror watching behaviors. Experimental results showed
that this method was not sensitive to the types and details of drivers' faces, and was robust to some
disturbance, and the overall detection accuracy rate was 96.1%.

Key words: traffic safety, vehicle vision, driver, outer contour of face and neck, rearview mirror
watching behavior