Accurate　detection　of　muscle　activity　is　crucial　for　rehabilitation　systems　that　rely　on　voluntary　control.　However,　the　presence　of　false　background　spikes　interference　can　significantly　impede　the　precise　decoding　of　motion　intention.　To　address　this　problem,　an　adaptive　two-step　method　is　proposed　in　this　paper　for　accurately　extracting　the　muscle　activation　intervals　from　sEMG　signals.　In　the　first　step,　an　adaptive　threshold　is　used　to　identify　potential　onsets　and　offsets　of　the　envelope.　Then,　a　combination　of　an　evaluation　equation　and　the　k-means　clustering　technique　is　utilized　to　eliminate　incorrect　onsets　and　offsets.　In　the　second　step,　two　peak　points　closing　proximity　to　the　onset　and　offset　are　identified.　The　tangency　of　the　envelope＇s　onset　and　offset　with　the　respective　peak　points　is　then　determined,　with　the　intersection　point　of　these　tangencies　is　considered　as　the　final　onset　and　offset.　The　proposed　method　is　tested　on　semi-synthetic　sEMG　signals　and　real　sEMG　signals,　and　compared　with　state-of-the-art　algorithms.　The　results　clearly　indicate　that　the　proposed　method　produces　the　best　detection　performance,　and　eliminates　the　requirement　for　parameter　selection,　greatly　facilitating　the　signal　extraction　process.