Physical　posture　is　a　reflection　of　the　orderly　arrangement　of　the　body’s　bones　and　the　proper　functioning　of　its　muscle　tissue,　and　it　is　also　a　guarantee　of　good　health.　Evaluating　physical　posture　generally　involves　detecting　key　points　on　the　human　body,　which　is　essentially　a　dense　detection　task　in　machine　vision.　This　paper　proposes　a　new　key　point　detection　method　for　evaluating　physical　posture,　based　on　multi-scale　feature　fusion　and　self-attention　mechanism.　The　self-attention　mechanism　is　added　in　the　algorithm　to　capture　global　feature　dependencies,　while　the　patching　merging　down-sampling　structures　help　reduce　information　loss　during　the　down-sampling　process.　Additionally,　a　de-convolution　module　is　added　in　the　prediction　phase　to　generate　higher　quality　feature　maps　and　improve　the　spatial　accuracy　of　the　key　points.　The　proposed　algorithm　achieves　an　average　mAP　of　85.5%　on　key　point　detection　models　for　the　front,　side,　and　back　of　the　body　on　a　self-built　dataset.　The　results　demonstrate　the　algorithm’s　good　performance　in　the　field　of　physical　posture　health　evaluation.　©　The　Author(s),　under　exclusive　license　to　Springer　Nature　Switzerland　AG　2023.