基于自适应特征感知的轻量化人体姿态估计  

Lightweight human pose estimation based on adaptive feature sensing

作　　者：毋宁 王鹏[2] 李晓艳[3] 吕志刚[3] 孙梦宇 WU Ning;WANG Peng;LI Xiao-yan;LÜZhi-gang;SUN Meng-yu(College of Ordnance Science and Technology,Xi'an Technological University,Xi'an 710021,China;Development Planning Office,Xi'an Technological University,Xi'an 710021,China;College of Electronics and Information Engineering,Xi'an Technological University,Xi'an 710021,China;College of Optoelectronic Engineering,Xi'an Technological University,Xi'an 710021,China)

机构地区：[1]西安工业大学兵器科学与技术学院,陕西西安710021 [2]西安工业大学发展规划处,陕西西安710021 [3]西安工业大学电子信息工程学院,陕西西安710021 [4]西安工业大学光电工程学院,陕西西安710021

出　　处：《液晶与显示》2023年第8期1107-1117,共11页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然基金(No.62171360);陕西省科技厅重点研发计划(No.2022GY-110);西安工业大学校长基金面上培育项目(No.XGPY200217);西安市智能兵器重点实验室(No.2019220514SYS020CG042)。

摘　　要：针对现有人体姿态估计网络在追求高精度检测时,网络结构设计复杂、模型参数量较大、检测效率较低的问题,本文提出了一种基于自适应特征感知的轻量级人体姿态估计算法。首先利用轻量化Ghost模块重构人体姿态估计的特征提取网络,减少网络参数量;其次设计了一种轻量级自适应特征感知的注意力机制,在降低网络模型复杂度的同时加强通道间信息有效交流,有效改善关键点定位效果;最后采用Huber Loss损失函数优化模型训练,实现异常点的更优预测,增强模型鲁棒性。在COCO数据集上进行验证,实验结果表明,与基准RMPE算法相比,改进后模型的检测精度提升了约0.5%,参数量减少了56.0%,网络运算量降低了32.6%,模型体积压缩了约57.0%,模型检测速率提升约2.1倍。本文改进后的人体姿态估计模型在压缩模型体积的同时提高了检测效率,增强了模型鲁棒性。For the problems of complex network structure design,large number of model parameters and low detection efficiency in the existing human pose estimation network pursues high-precision detection,this paper proposes a lightweight human pose estimation algorithm based on adaptive feature perception.Firstly,the lightweight Ghost module is used to reconstruct the feature extraction network of human pose estimation to reduce the computation amount of the network.Secondly,a lightweight adaptive feature sensing attention mechanism is designed to reduce the complexity of network model and enhance the effective communication between channels,which can improve the positioning effect of key points.Finally,Huber Loss(Exponential square loss function)is used to optimize the loss function training model to achieve better prediction of outliers and enhance the robustness of the model.Verified on the COCO dataset,the experimental results show that compared with the benchmark RMPE algorithm,the detection accuracy of the improved model is increased by about 0.5%,the number of parameters is reduced by 56.0%,the network calculation amount is reduced by 32.6%,the model volume is compressed by about 57.0%,and the model detection rate is increased by about 2.1 times.In this paper,the improved human pose estimation model improves the detection efficiency and enhances the robustness of the model while compressing the model volume.

关 键 词：人体姿态估计 轻量化 自适应特征感知 Ghost模块 Huber Loss 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]