基于人体关键点的分心驾驶行为识别-《计算机技术与发展》

文章信息/Info

Author(s):: XIA Han-sheng; SHEN Huan; HU Wei; School of Energy and Power Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China

Keywords:: driver distraction; human pose keypoints; convolutional neural network; heat maps; deep learning

摘要:: 驾驶员分心驾驶是造成交通事故的主要原因之一,利用车载设备识别驾驶员是否存在分心行为是当下亟须解决的问题。识别驾驶员是否存在分心行为的关键,在于正确理解驾驶员的姿态。对此,文中提出一种使用驾驶员的人体关键点位置信息来帮助卷积神经网络识别驾驶员是否分心驾驶的方法。通过加入人体关键点的位置信息,可以有效地使得卷积神经网络关注于驾驶员的姿态,减少背景信息的干扰。使用 Alpha Pose 系统获取驾驶员上半身9 个关键点的坐标,利用高斯公式分别以每个关键点为中心生成热力图。热力图包含关键点位置的响应,离关键点越近的位置,响应值越大。在 VGG16 和 ResNet50 的基础上,探讨8 种结构,分别将9 张热力图和不同的特征图融合,作为下一个卷积的输入。实验结果表明,该方法在 State Farm 数据集上达到了 94.934%的准确率,优于其他方法。

Abstract:: Detecting distraction of drivers is one of the main causes of traffic accidents. Using in-vehicle equipment to identify whether the driver has distracted behavior is an urgent problem to be solved. The key to identify whether the driver has distracted behavior is to correctly understand the driver’s posture. For this,we propose a method to help the convolutional neural network identify whether the driver is distracted by driving by human keypoints. By adding the position information of human keypoints,the convolutional neural network can effectively focus on the driver’s attitude and reduce the interference of background information. The Alpha Pose system is used to obtain the coordinates of 9 keypoints of the driver’s upper body,and Gauss formula is used to generate the heat map with each keypoint as the center. The heat map contains the response of the keypoints. The closer to the keypoints,the higher the response value. On the basis of VGG16 and ResNet50,8 structures are discussed,and 9 heat maps and different characteristic graphs are respectively fused as the input of the next convolution. The experiment shows that the proposed method has an accuracy rate of 94.934% in the State Farm Dataset,which is better than other methods.