基于姿态估计的驾驶员手部动作检测研究

为有效检查驾驶员在行驶过程中的不当行为，本文研究结合人体姿态估计信息的检测算法，通过对检测目标的约束，建立起一套具有多阶段的手部动作检测方法。该方法包含三个模块。第一，人体姿态估计模块，选取人体姿态估计网络关节的高斯热图层，通过输出的人体姿态高斯热图信息，达到对检测目标的空间信息的获取；第二，手部检测模块，基于CNN的检测网络，在网络输入层融合人体姿态高斯热图后，达到对手部的检测率提高的效果；第三，手部动作分类模块，通过接受手部检测模块的输出，消除对检测结果产生干扰的背景，将分类网络的特征提取约束在手部局部位置，提高手部动作分类的准确率，将手部区域输入至分类网络得到驾驶员手部动作，从而判断驾驶员是否存在抽烟、接听电话等不当行为，实现驾驶员的行为检测。为了验证本文提出的多阶段的手部动作检测方法，已在自制数据集上进行了相应实验。 

Driver's Hand Motion Detection Based on Pose Estimation

In order to effectively detect bus drivers’ misbehavior. Paper proposed an algorithm that can detect the irregular behavior of bus drivers while driving, and studied the detection algorithm combined with human pose estimation information for improving the accuracy of detection in the target environment. Finally, a multi-stage hand motion detection method is established. The method consists of three modules. Firstly, human pose estimation module, The Gaussian heat map layer of the human pose estimation network is selected, and the Gaussian heat map information of the human pose is output to achieve the constraint on the detection target space information. Secondly, the hand detection module, based on the CNN detection network, achieves the effect of improving the detection rate of the hand after the human pose Gaussian heat map is integrated at the network input layer. Thirdly, hand motion classification module, by accepting the output of the hand detection module, eliminating the background of interference with the detection result, constraining the feature extraction of the classification network to the local area of the hand, and improving the accuracy of the hand motion classification to the final Human body motion detection effect. In order to verify the multi-stage hand motion detection method proposed in this paper, corresponding experiments have been performed on homemade data sets.