多角度视频的驾驶人员行为识别

驾驶员在行驶过程中看手机，与乘车人员交谈等违规行为，为安全行驶造成了极大的隐患。为了解决此 类问题，提出了一种多角度行为识别方法，从3 个角度同步捕捉驾驶员行为的视频，构建多角度驾驶员行为的 视频和数据集，利用深度卷积神经网络，进行识别分类。实验结果表明，3D CNN 相对于2D CNN 的识别精度 更加准确，在对比输入剪辑的帧数实验中，发现堆叠的视频帧数会影响准确度，并在具有较大优势的R2plus1D 模型中( 将3D 卷积滤波器分解为单独的空间和时间分量) ，基于多角度驾驶人员行为识别精度达到87%。

Research on Driver Action Recognition Based on Multi-Angle Videos

In the process of driving，violations such as looking at mobile phones and talking with passengers cause great hidden dangers to safe driving． In order to solve such problems，we propose a multi-angle behavior recognition method，which can synchronously capture video of driving behaviors from three angles，construct a multi-angle driver behavior data set，and use deep convolutional neural network to carry out recognition and classification． The experimental results show the advantages of 3D CNN over 2D CNN in accuracy． And in comparing input clip frames experiment，we found the stacked video frames will affect accuracy and in R2plus1D model ( 3D convolution filter is decomposed into a separate component of time and space) ，realized driver behavior based recognition accuracy of 87% from multiple perspectives．