基于驾驶员视线注视和转移行为的驾驶分心状态检测

With the process of urbanization and the rapid increase of the number of motor vehicles, China's traffic safety situation is facing severe challenges. The dangerous driving behavior of drivers is the major cause of road traffic accidents, including fatigue driving, distraction driving, anger driving, and drunk driving. Due to the continuous increase of distraction sources and difficulties in attention monitoring, distraction driving occurs at the highest frequency which has become a serious threat to road safety. In this project, video images of the driver's face are first obtained through the on-board camera. To estimate and track the driver's gaze direction, a joint estimation algorithm based on deep convolution neural network is proposed for simultaneous landmarks localization, pose estimation, relative gaze direction estimation and absolute gaze direction estimation. It exploits the synergy among the tasks which boosts up their individual performances under the actual driving environment. Then an adaptive method for gaze region division is proposed to overcome the difference of individual drivers and vehicle type, which can accurately divide gaze region corresponding to each focus area. Finally, a robust and accurate driver distraction detection method is proposed based on support vector machine (SVM), where the feature space is generated using driver's gaze fixation histogram and diversion histogram, to alarm and intervene the dangerous state of drivers. The study has important theoretical significance and social application value for reducing the incidence of traffic accidents, improving the safety of driving behavior and improving the safety of road traffic in China.

伴随着城市化的进程及机动车数量的剧增，我国的交通安全形势面临严峻的挑战。驾驶员的危险驾驶行为是道路交通事故的主要致因，包括疲劳驾驶、分心驾驶、愤怒驾驶等，因驾驶员面对的分心源持续增长和注意力监测困难等原因，致使分心驾驶发生频率最高，成为当今道路安全的严重威胁。本项目通过车载摄像机获取驾驶员脸部视频图像，对其视线方向进行估计与跟踪，提出一种基于深度卷积神经网络的人脸特征点、头部姿态、相对视线方向和绝对视线方向的联合估计算法，提高实际行车环境下视线方向的估计精度；提出自适应视线注视区间划分方法，能克服驾驶员个体和车型差异影响，准确划分各注视区域所对应的视线区间；然后基于各区间的视线转移和注视直方图建立特征空间，采用支持向量机模型实现分心驾驶状态的高可靠性检测，从而对驾驶员进行报警干预。该研究对于降低交通事故的发生率，提高驾驶行为安全性和改善我国道路交通安全状况具有重要的理论意义和社会应用价值。

伴随着城市化的进程及机动车数量的剧增，我国的交通安全形势面临严峻的挑战。驾驶员的危险驾驶状态或行为是道路交通事故的主要致因，包括分心驾驶、疲劳驾驶、愤怒驾驶等，成为当今道路安全的严重威胁。驾驶员的危险驾驶行为检测的难点在于各种危险行为尤其是分心驾驶、疲劳驾驶等发生率高，检测困难，而现有检测方法相对简单、环境适应性差，无法满足各种复杂工况的检测需求。本项目通过车载摄像机获取驾驶员脸部视频图像，针对驾驶人状态检测过程中常存在人脸遮挡造成人脸检测准确率低、特征点检测精度差问题，提出基于ShuffleNet改进的MTCNN人脸检测模型算法，大幅提升人脸检测的效率；并提出了基于多任务学习的无需任何辅助标注信息的驾驶员人脸特征点检测方法，驾驶人的去人脸遮挡的人脸关键点检测算法，实现了人脸遮挡下的特征点检测；针对视线估计方法中存在双眼间远距离依赖关系捕获能力差、脸部图像冗余信息干扰视线预测等诸多造成视线估计精度不高问题，提出融合条纹池化和交叉注意力机制的视线估计模型SPMCCA-Net；在驾驶员注视分区相关的驾驶场景理解方面，提出了基于卷积神经网络与传统算法相结合的实时复杂道路车道线检测方法，大大提升了检测精度和检测速度，还提出了一种解决遮挡和目标尺寸差异大的城市道路全景分割方法，以及快速、准确地检测交通标志的方法。最后，利用特征点、视线方向等提取特征，建立驾驶员行为和状态检测模型，实现了复杂驾驶环境下驾驶人员状态检测与识别。驾驶人员状态检测与识别对于降低交通事故发生率，提高驾驶行为安全性和改善我国道路交通安全状况具有重要的理论意义和社会应用价值。