Deep neural networks (DNN) have recently demonstrated the ability to use big data to predict the traffic flow. However, the disadvantage of DNNs is that a large amount of data needs to be collected for each intersection and different intersections need to train different deep networks to estimate traffic flow accurately. This study proposes a new adaptive signal management method for the overall processing of smart cities, which combines deep learning and simulation to balance the issue of large-scale data collection. First, a computer-vision-based deep-learning network is trained offline to detect different types of vehicles. A large amount of training data can be collected throughout the city or country of interest, and the deep network only needs to be trained once. Then, for each intersection where traffic flow should be predicted, a small amount of data is collected, and a computer simulation model is developed to estimate local traffic flow. Finally, combining the traffic monitoring system based on deep learning with optimized simulation results, an adaptive traffic light management algorithm is developed. The proposed method can be easily adapted to different intersections by collecting a small amount of traffic data for each new intersection. Experimental results with real data for a complex T-shaped intersection in Macao show that the proposed method can significantly improve the overall traffic efficiency.