With　the　widespread　adoption　of　pose　estimation　technology　in　computer　vision,　traditional　methods　often　struggle　with　insufficient　accuracy　and　poor　robustness　when　dealing　with　complex　dynamic　scenes　and　diverse　backgrounds.　To　overcome　these　challenges,　we　propose　a　novel　pose　estimation　approach　that　integrates　a　dynamic　graph　convolutional　network　(DGCN),　a　spatiotemporal　interleaved　attention　mechanism　(STIA),　and　a　variable-length　Transformer　encoder　(VLTE).　The　DGCN　module　captures　the　spatial　dependencies　of　human　poses　and　enhances　the　model＇s　ability　to　represent　complex　spatial　relationships　by　dynamically　adjusting　the　graph　structure.　The　STIA　module　effectively　captures　spatiotemporal　dependencies　by　combining　both　temporal　and　spatial　information.　The　VLTE　module　improves　the　model＇s　adaptability　to　varying　time　scales　by　processing　variable-length　sequences　and　multi-scale　information.　We　performed　extensive　experimental　validation　on　the　MPI-INF-3DHP　dataset,　and　the　results　demonstrate　that　removing　any　of　the　modules　significantly　reduces　the　model＇s　performance,　highlighting　the　importance　of　each　component　in　the　overall　architecture.　Furthermore,　inference　speed　analysis　reveals　that　while　streamlining　the　model　can　enhance　inference　speed,　it　comes　at　the　cost　of　accuracy　and　robustness.　Our　work　presents　a　pose　estimation　solution　that　strikes　a　balance　between　computational　efficiency　and　high　accuracy,　offering　strong　support　for　practical　pose　estimation　tasks.　©　2025