Abstract:

This paper introduces the theory of face feature point detection, and proposes a solution based on Deep Convolutional Neural Network（DCNN） to solve the five facial feature points（eye, nose, mouth） prediction problem. By adding more convolutional layers to steadily increase the depth of the network, and using 3×3 convolutional filters in all layers, the parameters are effectively reduced, thus the problem of facial feature point detection is better solved. Then the homomorphism matrix of the plane formed by the corner of eyes and mouth and the plane when facing directly is calculated. Finally, the driver’s facial rotation angle is solved by the equivalent algorithm. Experimental results show that the detection accuracy of facial feature points is 97.96%, and the error of the algorithm in angle judgment is 1°-5°, which proves the effectiveness of the algorithm in the monitoring of distraction.

Key words: Deep Convolutional Neural Network（DCNN）, facial feature points detection, convolution and pooling, driver’s facial orientation

摘要：

在介绍人脸特征点检测的理论知识的基础上，提出了一种基于深层卷积神经网络（Deep Convolutional Neural Network，DCNN）解决人脸5点特征点（眼角、鼻子、嘴角）预测问题的方法。通过添加更多的卷积层稳定地增加网络的深度，并且在所有层中使用3×3的卷积滤波器，有效减小参数，更好地解决了人脸特征点检测问题。然后计算双眼角与嘴角所成平面与正视时此平面的单应性矩阵，最后利用等效算法求解驾驶员面部转角。实验结果表明，面部特征点检测准确率达到97.96%，算法在角度判断上的误差是1°~5°，这证明了该算法对注意力分散监测的有效性。

关键词: 深度卷积神经网络（DCNN）, 面部特征点检测, 卷积层和池化层, 驾驶员面部朝向