Vision-based　vehicle　behaviour　analysis　has　drawn　increasing　research　efforts　as　an　interesting　and　challenging　issue　in　recent　years.　Although　a　variety　of　approaches　have　been　taken　to　characterise　on-road　behaviour,　there　still　lacks　a　general　model　for　interpreting　the　behaviour　of　vehicles　on　the　road.　In　this　Letter,　the　authors　propose　a　new　method　that　effectively　predicts　the　vehicle　behaviour　based　on　structured　deep　forest　modelling.　Inspired　by　structured　learning,　the　structure　information　of　vehicle　behaviour　is　extracted　from　the　detected　vehicle,　and　then　the　corresponding　structured　label　is　constructed.　Especially,　the　structured　label　visually　expresses　the　vehicle　behaviour　as　contrast　to　the　discrete　numerical　label.　With　the　structured　label,　a　structured　deep　forest　model　is　proposed　to　predict　the　vehicle　behaviour.　Experimental　results　illustrate　that　the　proposed　method　successfully　obtains　the　implication　of　semantic　interpretation　of　vehicle　behaviour　by　the　predicted　structured　labels,　and　meanwhile　it　achieves　comparable　performance　with　traditional　methods.