Human　activity　recognition　(HAR)　has　been　a　vital　human–computer　interaction　service　in　smart　homes.　It　is　still　a　challenging　task　due　to　the　diversity　and　similarity　of　human　actions.　In　this　paper,　a　novel　hierarchical　deep　learning-based　methodology　equipped　with　low-cost　sensors　is　proposed　for　high-accuracy　device-free　human　activity　recognition.　ESP8266,　as　the　sensing　hardware,　was　utilized　to　deploy　the　WiFi　sensor　network　and　collect　multi-dimensional　received　signal　strength　indicator　(RSSI)　records.　The　proposed　learning　model　presents　a　coarse-to-fine　hierarchical　classification　framework　with　two-level　perception　modules.　In　the　coarse-level　stage,　twelve　statistical　features　of　time–frequency　domains　were　extracted　from　the　RSSI　measurements　filtered　by　a　butterworth　low-pass　filter,　and　a　support　vector　machine　(SVM)　model　was　employed　to　quickly　recognize　the　basic　human　activities　by　classifying　the　signal　statistical　features.　In　the　fine-level　stage,　the　gated　recurrent　unit　(GRU),　a　representative　type　of　recurrent　neural　network　(RNN),　was　applied　to　address　issues　of　the　confused　recognition　of　similar　activities.　The　GRU　model　can　realize　automatic　multi-level　feature　extraction　from　the　RSSI　measurements　and　accurately　discriminate　the　similar　activities.　The　experimental　results　show　that　the　proposed　approach　achieved　recognition　accuracies　of　96.45%　and　94.59%　for　six　types　of　activities　in　two　different　environments　and　performed　better　compared　the　traditional　pattern-based　methods.　The　proposed　hierarchical　learning　method　provides　a　low-cost　sensor-based　HAR　framework　to　enhance　the　recognition　accuracy　and　modeling　efficiency.　©　2021　by　the　authors.　Licensee　MDPI,　Basel,　Switzerland.