With the development of the city, the number of vehicles is increasing. This phenomenon not only increases urban congestion, but also leads to frequent traffic accidents. In order to improve the ability of urban governance, it is necessary to improve the monitoring ability of urban vehicles. In this paper, UAV is used to take low-altitude photography of four scenes in Shanghai and Chifeng area, and aerial remote sensing image data are obtained. Then, single target extraction is carried out for vehicles in UAV images combined with Unet convolutional neural network technology of deep learning. The results show that the ability of deep learning to recognize vehicles in UAV images is much higher than the random forest method in traditional machine learning, which reaches an ultra-high accuracy of 99%. And the estimation result of the number of cars in each scene is very close to the real number. According to the research results, the vehicle detection method combining UAV and deep learning technology has real time and practical feasibility, which can provide reliable technical means for real-time vehicle monitoring and traffic management in cities.