基于LHPN算法的手势姿态估计方法研究  

Research on Hand Pose Estimation Using LHPN Algorithm

作　　者：周全[1] 甘屹[1,2] 何伟铭[1,2] 孙福佳[1] 杨丽红[1] ZHOU Quan;GAN Yi;HE Wei-ming;SUN Fu-jia;YANG Li-hong(College of Mechanical Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Department of Precision Mechanics,Faculty of Science and Engineering,Chuo University,Tokyo〒112-8551,Japan)

机构地区：[1]上海理工大学机械工程学院,上海200093 [2]日本中央大学,理工学部精密工学科,日本东京112-8551

出　　处：《软件》2020年第7期66-71,共6页Software

基　　金：国家自然科学基金项目(批准号:51375314)。

摘　　要：随着广大用户越来越追求人工智能产品的体验,手势姿态估计存在广阔的应用前景,但也是当今计算机视觉的难题。针对目前自顶向下的姿态估计模式容易受视觉传感器与目标检测精度的影响,本文提出基于轻量级手势姿态网络(LightweightHandPoseNet,LHPN)算法的手势姿态估计方法,该算法采用ConvolutionalPose Machines(CPM)算法的多层次顺序结构,在每个阶段后隐式地将上下文信息融合,并设计了轻量级主干网络,以提升手势姿态估计的综合性能。基于STB数据集对比分析不同内部结构的LHPN算法性能,并与典型算法进行对比。实验结果表明,LHPN算法能够对手势姿态进行准确估计,与CPM算法相比,在AUC方面提升了0.5%,在每帧图像运算时长方面减少了0.1358 s。As users increasingly pursue the experience of artificial intelligence products, gesture pose estimation has broad application prospects, but it is also a difficult problem in computer vision. In view of the fact that the current top-down hand pose estimation mode is easily affected by visual sensors and object detection accuracy, this paper proposes a hand pose estimation method based on(Lightweight Hand Pose Net, LHPN) algorithm. The algorithm uses the multi-level sequence structure of Convolutional Pose Machines(CPM) algorithm, implicitly combines the context information after each stage, and designs a lightweight backbone network to improve the comprehensive performance of hand pose estimation. Based on STB dataset, the performance of LHPN algorithm with different internal structures is analyzed and compared with typical algorithms. The experimental results show that LHPN algorithm can accurately estimate hand pose. Compared with CPM algorithm, it improves AUC by 0.5% and reduces computation time of each frame of image by 0.1358 s.

关 键 词：手势姿态估计 Lightweight Hand Pose Net Convolutional Pose Machines 轻量级 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]