An　image　feature　extraction　technique　based　on　histogram　of　oriented　gradients　(HOG)　technology　is　proposed　as　a　method　for　human　body　detection,　while　3D　convolutional　neural　networks　(3D　CNN)　technology　is　combined　as　a　key　technology　for　action　recognition.　And　the　two　are　combined　to　generate　3DHOG　assistive　technology　which　is　applied　to　the　physical　education　video　parsing.　The　results　show　that　the　false　recognition　rate　of　3D　CNN　model　in　the　training　set　is　stable　around　0.03,　corresponding　to　a　loss　of　0.05.　The　average　accuracy　of　each　action　of　3D　HOG　model　is　96.25%,　while　the　recall　rate　of　the　model　is　81.2%.　Its　mean　absolute　error　(MAE)　value　is　1.18%　and　root　mean　squared　error　(RMSE)　value　is　0.91%.　The　3D　HOG　model　has　superior　performance　and　has　good　application　value　for　action　detection　and　recognition　of　physical　education　videos.　Copyright　©　2025　Inderscience　Enterprises　Ltd.