Application　of　human　pose　estimation　bring　great　help　to　people＇s　life.　Most　of　the　applications　in　real　life　scenes　are　based　on　single-frame　images.　The　work　based　on　a　single-frame　image　has　better　accuracy,　but　it　often　abandons　some　temporal　information　of　real　life.　In　order　to　preserve　the　information,　we　choose　to　combine　the　Long　Short-Term　Memory　Network　with　a　single-frame　estimation　network　to　carry　out　the　multi-person　pose　estimation　for　video　stream.　In　the　design　of　single-frame　network,　this　paper　adds　the　deconvolution　layers　to　the　residual　network　to　obtain　high-resolution　image　information　and　adds　a　loss　function　with　an　area　of　bounding-box　to　train　the　single-frame　model.　In　the　design　of　multi-person,　pose　estimation　network　for　video　Stream,　this　paper　uses　the　Long　Short-Term　Memory　Network　to　process　the　temporal　information　extracted　from　the　single-frame　network　to　carry　out　multi-person　pose　estimation.　In　the　experiment,　COCO　dataset　and　PoseTrack2018　dataset　verify　the　effectiveness　of　the　method.　©　2019　IEEE.