基于YOLOv3-HM的手部姿态估计方法研究

Research on Hand Pose Estimation Method Based on YOLOv3-HM

二维手部姿态估计是人机交互领域的一项关键技术。为增强复杂环境下系统鲁棒性,提高手势姿态估计精度,提出一种基于目标检测和热图回归的YOLOv3-HM算法。首先,利用YOLOv3算法从RGB图像中识别框选手部区域,采用CIoU作为边界框损失函数;然后,结合热图回归算法对手部的21个关键点进行标注;最终,通过回归手部热图实现二维手部姿态估计。分别在FreiHAND数据集与真实场景下进行测试,结果表明,该算法相较于传统手势检测算法在姿态估计精度和检测速度上均有所提高,对手部关键点的识别准确率达到99.28%,实时检测速度达到59 f/s,在复杂场景下均能精准实现手部姿态估计。

2D hand pose estimation is a key technology in the field of human-computer interaction.In order to enhance the robustness of the system in complex environments and improve the accuracy of gesture estimation,a YOLOv3-HM algorithm based on target detection and heatmap regression is proposed.Firstly,the YOLOv3 algorithm is used to identify and frame hand region from RGB image,and CIoU is used as the loss function of the bounding box.Then,a heatmap regression algorithm is used to label the 21 key points of the hand.Finally,2D hand pose estimation is achieved by regression of the hand heatmap.The experiments are carried out on the FreiHAND dataset and the real scene respectively.The result show that the algorithm has improved the pose estimation accuracy and the detection speed compared with the traditional gesture detection algorithms.The recognition accuracy of key points of the hand reaches 99.28%,and the real-time detection speed reaches 59 f/s.The algorithm can achieve good hand pose estimation in different scenarios.