基于视频序列的危险驾驶预警技术研究

为减少因危险驾驶行为造成的交通事故，提出一种基于视觉的危险驾驶预警方法。依据人脸特征点定位结果计算嘴部几何特征值，并使用头部姿态角对几何特征进行修正，形成嘴部张开度和头部姿态的特征序列；针对数据丢失问题，使用滑动平均法进行缺失值补偿；统计分析时间窗内序列，对哈欠、说话、视线偏移行为进行识别；并提出一种多参数融合方案和非线性映射模型将危险可能性直观化，计算危险系数并进行危险驾驶报警。实验表明，哈欠、说话检测准确率分别达到82.97%、100%，头部姿态、危险驾驶系数与实际情况相符，模型效果好。

Research on Dangerous Driving Warning Based on Video Sequence

In order to reduce the traffic accidents caused by dangerous driving behaviors,a vision-based dangerous driving warning method is proposed.According to the facial feature point positioning results,the geometric feature value was calculated,and corrected by the angle of head pose.On the foundation of opening degree and the head pose,the time series were established.The missing data was compensated by sliding average method.The sequence within the time window was statistically analyzed to recognize the behaviors such as yawn,speech,and line-of-sight migration.A multi-parameter fusion scheme and nonlinear mapping model were put forward to visualize the possibility of danger.The risk factor was calculated for dangerous driving alarms.Experiments show that accuracy rates of yawn and speech detection reach 82.97% and 100% respectively,and the head pose and driving risk coefficient are consistent with the actual situation.The system performs well in realtime test.