基于单幅图像Pn P头部姿态估计的学习注意力可视化分析

Visualization analysis of learning attention based on single-image PnP head posture estimation

提出了一种基于单幅图像头部姿态估计的学生注意力可视化分析方法,采用随机级联回归树进行人脸特征点定位,引入了一个统计测量获得的刚性模型作为3D人脸近似,实现基于Pn P (perspective-n-point)映射的单幅图像头部姿态估计,最后将学生视线投射到教师授课的视频图像上,实现学生学习注意力的可视化分析。实验结果表明:对于Biwi标准库,该方法可以将头部姿态估计角度平均误差降低到4.88°;方法具有粗颗粒度的计算并行性,使用4线程并行计算可以获得2.37倍的加速效果;实现了3种典型学习状态(专注、关注、漠视)的注意力可视化分析。

Owing to the fact that the head-mounted eye tracker is unsuitable to be widely used in the large-scale classroom evaluation under expenditure limitation,a novel method was proposed for student learning attention visualization analysis based on single-image PnP head posture estimation.The method applicate the stochastic cascade regression tree technique to locate facial feature points.With the feature points,the PnP(perspective-n-point)mapping technique is used to estimate the head posture based on a 3D rigid facial model obtained through statistical measurement.The method finally projects the gaze point on the frame image of the teaching video,which realizes the visualization of student learning attention.Experiments demonstrate the following advantages of the method.1)The method limits the average head-posture estimation errors under 4.88°with Biwi database.2)Thanks to the coarse-grained computing parallelism of the method,the work achieves 2.37X speedup with four threads.3)the work has implemented student learning attention visualization analyses for three typical learning cases including engagement,attention,disregard.