Surface　electromyography　(sEMG)　signals　demonstrate　how　the　muscles　react　to　the　control　strategies　of　the　nervous　system.　For　stroke　patients,　the　detection　of　muscle　activation　is　crucial　because　it　allows　researchers　to　better　understand　their　brain　control　mechanisms.　However,　extracting　the　muscular　activity　precisely　from　the　sEMG　signals　with　various　complex　noises　is　a　challenge　in　biomedical　data　processing.　This　study　presented　an　adaptive　two-step　(AdaTS)　muscular　activity　recognition　algorithm　to　handle　this　problem.　The　sEMG　signal　was　first　pre-extracted　using　the　adaptive　threshold　approach　and　annexation　method　to　identify　the　interval　comprising　active　information.　Then,　after　amplifying　the　difference　of　activity　and　inactivity　of　the　interval　signal　through　the　Teager-Kaiser　energy　operator,　the　onsets　and　offsets　were　determined　based　on　the　overall　change　of　the　interval　signal.　The　proposed　algorithm　was　tested　in　semi-synthetic　signals,　real　signals　from　a　public　database,　and　experimentally　recorded　signals.　Compared　with　other　approaches,　our　method　can　effectively　handle　various　types　of　interference　and　produce　the　best　detection　performance.　Additionally,　the　steps　of　parameter　selection　and　adjustment　are　removed,　which　greatly　simplifies　the　practical　application.