It　is　a　commonly　known　fact　that　both　alcohol　and　fatigue　impair　driving　performance.　Therefore,　the　identification　of　fatigue　and　drinking　status　is　very　important.　In　this　study,　each　of　the　22　participants　finished　five　driving　tests　in　total.　The　control　condition,　serving　as　the　benchmark　in　the　five　driving　tests,　refers　to　alert　driving.　The　other　four　test　conditions　include　driving　with　three　blood　alcohol　content　(BAC)　levels　(0.02%,　0.05%,　and　0.08%)　and　driving　in　a　fatigued　state.　The　driving　scenario　included　straight　and　curved　roads.　The　straight　roads　connected　the　curved　ones　with　radii　of　200　m,　500　m,　and　800　m　with　two　turning　directions　(left　and　right).　Driving　performance　indicators　such　as　the　average　and　standard　deviation　of　longitudinal　speed　and　lane　position　were　selected　to　identify　drunk　driving　and　fatigued　driving.　In　the　process　of　identification,　road　geometry　(straight　segments,　radius,　and　direction　of　curves)　was　also　taken　into　account.　Alert　vs.　abnormal　and　fatigued　vs.　drunk　driving　with　various　BAC　levels　were　analyzed　separately　using　the　Classification　and　Regression　Tree　(CART)　model,　and　the　significance　of　the　variables　on　the　binary　response　variable　was　determined.　The　results　showed　that　the　decision　tree　could　be　used　to　distinguish　normal　driving　from　abnormal　driving,　fatigued　driving,　and　drunk　driving　based　on　the　indexes　of　vehicle　speed　and　lane　position　at　curves　with　different　radii.　The　overall　accuracy　of　classification　of　alert　and　abnormal　driving　was　90.9%,　and　that　of　fatigued　and　drunk　driving　was　94.4%.　The　accuracy　was　relatively　low　in　identifying　different　BAC　degrees.　This　experiment　is　designed　to　provide　a　reference　for　detecting　dangerous　driving　states.