一种基于深度学习的驾驶员打电话行为检测方法

A Driver's Calling Behavior Detection Method Based on Deep Learning

针对现有驾驶员打电话行为检测方法存在精度低、实时性能差以及免提通话无法检测等问题,提出一种基于深度学习的多特征融合检测方法。该方法通过融合手持电话行为与讲话行为检测结果,实现对驾驶员打电话行为的检测。基于构建的浅层卷积神经网络,包含5层特征提取网络以及2层全连接层,可实现对听筒及免提接听两种手持电话行为的检测;同时,通过级联形状回归算法得到嘴部18个特征点及其宽高比,并根据连续20帧图像的嘴巴宽高比振荡差值来检测讲话行为。实验结果表明,该方法在实际驾驶场景下的平均检测准确率达到95.6%,平均检测耗时低至230 ms/frame,综合检测性能得到明显改善。

For the problems of low accuracy,poor real-time performance and inability to detect hands-free calls in the existing detection methods of drivers calling behavior,a multi-feature fusion detection method based on deep learning is proposed.The method combines the results of hand holding behavior and speech behavior to detect the Driver's calling behavior.The constructed shallow convolutional neural network(S-CNN)consists of five feature extraction layers and two fully-connected layers.It can detect two types of handset behavior:telephone receiver and hands-free answering.At the same time,the speech behavior is detected by the oscillation difference of the mouth width to height ratio of 20 consecutive images,which is obtained by the 18 feature points of the mouth generated by the cascading shape regression algorithm.The experimental results show that the average detection accuracy of the proposed method in actual driving scenarios reaches 95.6%,and the average detection time is as low as 230 ms/frame.The comprehensive detection performance of the method is improved obviously.