Provided is an illegal-parking vehicle and breakdown vehicle vision detection system based on deep convolution nerve network, comprising a camera disposed on a city road, a traffic cloud server, and aroad traffic event automatic detection system. In the system, various vehicles on the road can be extracted through deep convolution nerve network technology, and the optical flow method is adopted to calculate, identify and determine whether the vehicles are still or not. If a still vehicle is identified, and the rest time exceeds the parking time threshold value, the vehicle will be determinedas illegal parking. Means like WebGIS or broadcasting and road caution boards will be adopted for real-time report, so that the traffic police will rapidly arrange people to remove the traffic obstacles, and simultaneously remind the following vehicles of the accidents on the road in front of them. Accordingly, relevant measures will be taken to avoid secondary accident. The invention is advantageous in that robustness is quite good, and identification precision is quite high.