摘要
相较于2017年提出的在当时检测效果近乎最优的RMPE模型与Mask R-CNN模型,原用于人体骨骼关键点检测的OpenPose模型有着在保持精度近乎不变的情况下能大幅缩短检测周期的优势,但同时该模型也存在着参数共享率低、冗余度高、耗时长、模型规模太大等问题。针对上述问题,提出了新的OpenPose-slim模型。该模型减小网络宽度,减少卷积块层数,将原并列式结构改成序列式结构并于内模块加入Dense连接机制,其处理过程主要分为3个模块:1)关键点定位模块,检测出人体骨骼关键点的位置坐标;2)关键点联系模块,把关键点位置连接成肢体;3)肢体匹配模块,进行肢体匹配得到人体轮廓。每一个处理阶段之间关联紧密。在MPII数据集、COCO数据集和AI Challenger数据集上的实验结果表明,所提模型使用4个定位模块和2个联系模块,并于每一个模块内部使用Dense连接机制是最佳结构,与OpenPose模型相比,在保持检测精度基本不变的基础上,测试周期缩短为原来的近1/6,参数量缩小了近50%,模型规模缩小为近1/27。
The OpenPose model originally used for the detection of key points in human skeleton can greatly shorten the detection cycle while maintaining the accuracy of the Regional Multi-Person Pose Estimation(RMPE) model and the Mask Region-based Convolutional Neural Network(R-CNN) model, which were proposed in 2017 and had the near-optimal detection effect at that time. At the same time, the OpenPose model has the problems such as low parameter sharing rate, high redundancy, long time-consuming and too large model scale. In order to solve the problems, a new OpenPose-slim model was proposed. In the proposed model, the network width was reduced, the number of convolution block layers was decreased, the original parallel structure was changed into sequential structure and the Dense connection mechanism was added to the inner module. The processing process was mainly divided into three modules: 1) the position coordinates of human skeleton key points were detected in the key point localization module;2) the key point positions were connected to the limb in the key point association module;3) limb matching was performed to obtain the contour of human body in the limb matching module. There is a close correlation between processing stages. The experimental results on the MPII dataset, Common Objects in COntext(COCO) dataset and AI Challenger dataset show that, the use of four localization modules and two association modules as well as the use of Dense connection mechanism inside each module of the proposed model is the best structure. Compared with the OpenPose model, the test cycle of the proposed model is shortened to nearly 1/6, the parameter size is reduced by nearly 50%, and the model size is reduced to nearly 1/27.
作者
汪检兵
李俊
WANG Jianbing;LI Jun(College of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan Hubei 430065,China;Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System(Wuhan University of Science and Technology),Wuhan Hubei 430065,China)
出处
《计算机应用》
CSCD
北大核心
2019年第12期3503-3509,共7页
journal of Computer Applications
基金
国家自然科学基金资助项目(61572381)
武汉科技大学智能信息处理与实时工业系统湖北省重点实验室基金资助项目(znxx2018QN06)~~