Keyframe　extraction　is　important　for　video　retrieval.　In　order　to　realize　frequency　adaptive　human　motion　sequence　resampling　and　achieve　high　quality　keyframe,　we　propose　a　new　keyframe　extraction　method　for　human　motion　sequence.　First,　we　define　the　inter-frame　similarity　metric　based　on　the　features　of　human　body　parts.　Then,　the　keyframe　extraction　is　realized　by　the　affine　propagation　clustering　algorithm.　The　proposed　method　starts　from　the　information　distribution　of　the　video　itself,　adaptively　searches　for　the　optimal　keyframe　of　the　video,　and　the　operation　speed　is　fast.　Finally,　the　evaluation　of　the　sequence　reconstruction　based　on　keyframe　is　verified.　A　comparative　experiment　conducted　on　the　CMU　database　verified　the　efficiency　of　our　method.