驾驶人注意力分散的图像检测与分级预警  

Driver Attention Distraction Detection and Hierarchical Prewarning Based on Machine Vision

作　　者：程文冬[1,2] 付锐[1] 袁伟[1] 刘卓凡[1] 张名芳[1] 刘通[1] 

机构地区：[1]长安大学汽车学院,西安710064 [2]西安工业大学机电工程学院,西安710032

出　　处：《计算机辅助设计与图形学学报》2016年第8期1287-1296,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：长江学者和创新团队发展计划项目(IRT1286);国家自然科学基金(61374196;61473046);陕西省教育厅专项科研计划项目(16JK1375);西安工业大学校长基金(XAGDXJJ15006)

摘　　要：针对当前驾驶人注意力分散(DAD)图像检测的研究中,眼睛、嘴唇等目标易受到光照与遮挡的干扰,头部姿态模型的鲁棒性与准确性不易保证的问题,提出基于鼻孔图像识别的注意力区域识别方法与DAD层级预警.首先建立基于BF-SSR光照均衡法的Adaboost-肤色模型来识别驾驶人脸区域,在此范围内根据色度、面积与圆度的聚类特征来检测鼻孔,依据成像面上的鼻孔坐标变化来建立头部俯仰与横摆姿态模型,并解决头部平动时的参数初始化问题;然后定义头部横摆角、俯仰角、鼻孔中心坐标偏移量作为特征向量集,建立注意力区域的SVM分类模型;最后根据注意力偏离的时长、分配比例以及偏离的必要性建立DAD分级预警.实验结果表明,该方法对光照、眼镜、头部运动等干扰的鲁棒性好,头部横摆与俯仰姿态的平均误差为5.5°和4.9°,SVM对驾驶人注意力区域的分类准确率为85.8%,DAD预警准确率为85.4%.Machine vision is the main method for driver attention distraction（DAD） detection. In the current researches eyes, lips and other targets would be easily interfered by light and occlusion. Moreover, recent head pose models are difficult to satisfy the robustness and accuracy of DAD detection. For this problem a DAD detection method and hierarchical prewarning is proposed based on nostril recognition in this paper. Firstly driver face region is detected by a fusion algorithm combined with Adaboost and adaptive skin model, which is pretreated by BF-SSR illumination equalization. Nostrils are then detected within face region according to the cluster characteristics of color, area and roundness. After that head pose model of yaw and pitch is set up according to nostril coordinates on imaging plane. Meanwhile the problem of model parameter initialization is solved during head translational motion. On the basis of the above, attention classification model is established by support vector machine（SVM）, which is trained by head rotation angles and nostril coordinate offsets. Finally hierarchical DAD prewarning method is set up according to the time length, distribution proportion and necessary level of visual attention diversion. Experimental results demonstrate that the method has robust adaptability to illumination, eyeglasses and head rotation. The average angle errors of head yaw and pitch are 5.5° and 4.9°, respectively. The classification accuracy rate of attention regions by SVM classifier reaches 85.8% and DAD prewarning accuracy reaches 85.4%.

关 键 词：驾驶人注意力分散 辅助驾驶系统 机器视觉 层级预警 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]