The　Fine-grained　Vehicle　recognition　is　easily　affected　by　small　visual　changes.　The　existing　recognition　methods　have　less　robustness　to　these　conditions　(such　as　illumination,　weather　changes,　etc.)　and　the　accuracy　of　vehicle　recognition　in　complex　environments　cannot　achieve　a　satisfying　result.　In　this　paper,　a　high-accuracy　fine-grained　vehicle　recognition　method　using　Convolutional　Neural　Network　for　urban　surveillance　videos　is　proposed.　Differing　from　the　conventional　CNN　based　methods,　which　usually　obtain　a　single　classification　model,　the　proposed　method　combines　Pre-training　and　hierarchical　fine-tuning　to　obtain　different　classification　model　that　can　adapt　to　the　change　of　illumination　conditions.　In　the　recognition　phase,　the　input　image　is　evaluated　firstly　using　the　proposed　image　quality　assessment　method.　Then,　the　classification　model　is　selected　adaptively　according　to　the　results　of　quality　evaluation　and　used　for　recognition.　To　test　our　method,　we　use　a　publicly　vehicle　recognition　dataset　for　Urban　Surveillance　Videos　provide　by　Yang.　Experimental　results　show　that　the　accuracy　of　the　proposed　method　is　up　to　98.89%,　which　is　0.5%　higher　than　Yang＇s　method.