Gesture　recognition　is　playing　a　vital　role　in　human-computer　interaction　and　smart　home　automation.　Conventional　gesture　recognition　systems　require　either　dedicated　extra　infrastructure　to　be　deployed　or　users　to　carry　wearable　and　mobile　devices,　which　is　high-cost　and　intrusive　for　large-scale　implementation.　In　this　paper,　we　propose　FreeGesture,　a　device-free　gesture　recognition　scheme　that　can　automatically　identify　common　gestures　via　deep　learning　using　only　commodity　WiFi-enabled　IoT　devices.　A　novel　OpenWrt-based　IoT　platform　is　developed　so　that　the　fine-grained　Channel　State　Information　(CSI)　measurements　can　be　obtained　directly　from　IoT　devices.　Since　these　measurements　are　time-series　data,　we　consider　them　as　continuous　RF　images　and　construct　CSI　frames　with　both　amplitudes　and　phase　differences　as　features.　We　design　a　dedicated　convolutional　neural　network　(CNN)　to　uncover　the　discriminative　local　features　in　these　CSI　frames　and　construct　a　robust　classifier　for　gesture　recognition.　All　the　parameters　in　CNN　are　automatically　fine-tuned　end-to-end.　Experiments　are　conducted　in　a　typical　office　and　the　results　validate　that　FreeGesture　achieves　a　95.8%　gesture　recognition　accuracy.　©　2018　IEEE.