This　paper　introduces　a　new　method　for　real-time　gesture　recognition　using　infrared　thermal　images.　We　propose　a　lightweight　gesture　recognition　network　based　on　CNN,　called　TGNet　(ThermoGestureNet).　TGNet　enhances　recognition　accuracy　through　a　knowledge　distillation　strategy　and　improves　inference　speed　by　removing　redundant　convolutional　kernel　parameters　using　model　pruning　techniques.　Experimental　comparisons　on　embedded　devices　with　mainstream　lightweight　networks　show　that　TGNet　demonstrates　excellent　performance　in　both　recognition　speed　and　accuracy.