Gesture,　which　carries　a　large　amount　of　human　meanings　is　one　of　the　most　important　and　complicated　way　in　human-computer　interaction.　Representation　learning,　the　representation　of　complex　objects　and　their　eigenvalue　can　be　discovered　by　machine　learning.　It　can　have　an　objective　and　comprehensive　representation　of　experimental　sample.　This　paper　applies　representation　learning　to　the　feature　extraction　of　gestures　and　establishes　a　static　gesture　recognition　system.　The　system　uses　Leap　Motion　as　teaching　data　collector.　As　a　result,　the　process　of　gesture　segmentation　and　the　error　from　it　would　be　avoided.　In　order　to　reduce　model　training　and　recognition　time,　this　system　uses　support　vector　machine　algorithm,　an　algorithm　with　fast　response　sifted　and　high　recognition　rate,　as　a　classifier.　In　the　process　of　feature　extraction,　the　machine　learning　algorithm　itself　is　used　to　explore　the　representation　of　gestures　and　the　extraction　of　feature.　Therefore,　the　traditional　empirical　method　of　extraction　would　be　replaced　and　the　precision　and　stability　of　gesture　recognition　system　would　be　improved.　Experimental　results　show　that　compared　with　the　traditional　gesture　recognition　system,　the　improved　gesture　recognition　system　has　higher　recognition　rate,　accuracy　and　anti-interference　ability.