摘要
为了提高视频中人体姿态估计检测效果,在保留结构化信息的同时弥补下采样导致的空间分辨率的损失,增加视频中检测效率,本文结合时序信息提出了一种轻量级双路神经网络帧间信息推理的视频人体姿态估计方法.首先,基于最新的人体关键点检测网络训练一个基于该方法两路融合全卷积网络,一路选用金字塔全卷积网络,并选用采用轻量级Inverted residuals作为网络模块,另外一路保持分辨率大小不变以减少空间分辨率的损失,然后提出了一种利用帧间关键点信息建立时序模型,从而推理预测帧的关键点信息.本文在PoseTrack数据集中与最新的方法进行比较,关键点检测mAP提高1.3%,速度提升20%,关键点跟踪MOT提高2.8%,经过实验验证,本文算法可以保留结构化信息的同时有效弥补空间分辨率的损失并提高检测精度,同时提高了视频检测中的速度.
In order to improve the detection effect of human pose estimation in video,make up for the loss of spatial resolution caused by dow n sampling w hile retaining structured information,and increase the detection efficiency in video,this paper proposes a method of human pose estimation in video based on the frame information inference of lightw eight tw o-w ay neural netw ork combined w ith temporal information.First of all,based on the latest human key detection netw ork training,a tw o-w ay fusion full convolution netw ork based on this method is trained,one is pyramid full convolution netw ork,and the other is light-w eight inverted residuals as the netw ork module,and the other keeps the resolution size unchanged to reduce the loss of spatial resolution.Then,a time sequence module is proposed by using the information of key points betw een frames And then infer the key information of the prediction frame.In this paper,compared w ith the latest methods in Posetrack data set,the key point detection map is increased by 1.3%,the speed is increased by20%,and the key point tracking M OT is increased by 2.8%.Through experimental verification,the algorithm in this paper can retain structured information w hile effectively making up for the loss of spatial resolution and improving the detection accuracy,w hile improving the speed in video detection.
作者
陈昱昆
汪正祥
于莲芝
CHEN Yu-kun;WANG Zheng-xiang;YU Lian-zhi(School of Optical-Electrical and Computer Engineering,University,University of Shanghai for Science and Technology,Shanghai 200093,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2020年第10期2219-2224,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61603257)资助。
关键词
姿态估计
关键点检测
关键点跟踪
双路全卷积网络
轻量级
帧间信息
attitude estimation
key detection
key tracking
tw o-w ay full convolution netw ork
lightw eight
inter frame information
