For　some　professional　sports,　it　is　required　to　supervise　and　analyze　the　athletics　pose　in　training　of　athletes.　In　order　to　facilitate　the　browse　of　training　videos,　it＇s　necessary　to　extract　the　key　frames　from　training　videos.　In　this　paper,　we　propose　a　deep　key　frame　extraction　method　for　analyzing　sport　training　videos.　To　alleviate　the　bias　from　complex　background,　Fully　Convolutional　Networks　(FCN)　is　employed　firstly　to　extract　the　foreground　region　which　contains　the　athlete　and　barbell.　Then　over　the　extracted　foregrounds,　Convolutional　Neural　Networks　(CNN)　are　leveraged　to　estimate　the　pose　probability　of　each　frame　and　extract　the　key　frames　by　the　maximum　probability　on　each　pose.　The　experimental　results　demonstrate　that　the　proposed　method　achieves　good　performance　in　key　frame　extraction　of　sport　videos　comparing　method.