3D　human　pose　estimation　is　a　very　important　task　in　human-computer　interaction.　Most　of　state-of-art　methods　are　based　3D　convolutional　neural　from　voxelized　grid　or　2D　convolutional　neural　network　from　depth　images.　The　perspective　distortions　problems　may　badly　effect　the　performance　of　the　depth　image-based　methods.　The　degradation　problem　may　badly　effect　the　performance　of　3D　CNN-based　methods.　To　overcome　the　perspective　distortions　problem,　we　use　voxelized　gird　as　input　of　proposed　network.　The　output　of　proposed　network　is　a　heatmap　for　each　joint,　and　then　our　method　gets　3D　coordinate　of　joint　from　the　heatmap.　To　overcome　the　degradation　problem　and　gradient　vanishing　problem,　our　network　use　residual　blocks,　which　can　make　deeper　network　be　easier　to　be　trained.　The　attention　mechanism　can　help　our　method　focus　on　the　more　important　information.　We　evaluated　our　method　on　a　public　human　pose　dataset　and　a　human　pose　dataset　collected　by　ourself.　The　experimental　results　show　the　efficiency　of　our　method.　©　2021　IEEE.