In　this　paper,　we　propose　a　novel　descriptor　method　for　human　action　recognition　based　on　depth　video　sequences.　The　proposed　method　improves　the　flexibility　of　action　recognition　by　using　local　multiresolution　pyramids　in　feature　space.　In　feature　extraction,　we　extract　polynormals　of　different　scales　and　compose　new　pyramid-polynormals　to　express　the　multilayer　apparent　information　of　a　local　subcube,　improving　the　discrimination　of　the　descriptor.　Moreover,　we　also　present　a　novel　group　sparse　constraint　dictionary　learning　method　to　reduce　the　correlations　between　different　sub-dictionaries　and　obtain　a　sparse　dictionary　with　more　discriminative　ability;　we　then　use　sparse　low-level　coding　features　by　utilizing　the　learned　sparse　dictionary　and　applying　the　spatial　average　pool　and　temporal　maximum　pool　to　aggregate　the　sparse　coefficients　into　a　high-dimensional　feature.　The　feature　vectors　extracted　from　each　grid　are　then　concatenated　as　the　final　P-SNV　descriptor.　The　descriptor　can　effectively　preserve　local　multi-layer　apparent　information　of　human　actions　while　eliminating　similar　contents　contained　in　the　different　categories　of　action,　effectively　improving　the　recognition　rate.　Experimental　results　on　four　public　benchmark　datasets　demonstrate　that　our　algorithm　achieves　superior　performance　compared　to　the　state-of-the-art　algorithms.　©　2015　by　Binary　Information　Press.