With　the　rapid　development　of　urban　rail　transit,　traffic　safety　has　become　the　focus　of　attention　and　people　are　paying　increasing　attention　to　the　prevention　of　fatigue　driving.　＂Gesture　and　oral　instructions　of　urban　rail　traffic　drivers＇＇　is　operational　actions　of　drivers　written　in　the　Chinese　metro　operation　specification.　It　is　a　method　to　prevent　drivers　from　fatigue　driving　and　ensure　safety.　However,　there　is　a　lack　of　scientific　detection　methods.　We　combine　the　standard　traffic　operational　actions　with　fatigue　action　to　construct　a　fatigue　detection　system　that　is　suitable　for　the　urban　rail　transit　industry.　The　system　includes　a　dynamic　tracking　model　for　the　large-scale　operation　of　rail　transit　drivers　and　a　dualinput　action　discrimination　model　based　on　a　three-dimensional　convolutional　neural　network　(3DCNN).　The　model　sets　the　skipping　frame　and　continuous　frame　as　two　inputs　of　the　model,　and　extracts　five　channels　of　information　from　the　two　inputs.　Dual-input　multi-channel　information　enables　the　model　to　learn　not　only　the　spatial　and　temporal　information　of　the　entire　action,　but　also　the　subtle　changes　of　the　action.　First,　we　trained　and　validated　the　dual-input　model　based　on　a　3DCNN　using　the　open　dataset　KTH,　which　contains　several　variations.　Then,　the　model　trained　on　KTH　was　migrated　to　our　data　using　the　transfer　learning　method,　which　saved　training　time　and　achieves　an　accuracy　of　98.41%.　This　transfer　learning　scheme　can　also　be　applied　when　new　categories　are　encountered　in　practice.　Finally,　we　discussed　and　envisaged　the　future　optimization　of　the　system.