Due to the need to detect and track stolen and criminal vehicles and traffic monitoring, the development of license plate recognition systems on intersection monitors system is very urgent for the development of smart cities. Unlike the traditional license plate recognition technology applied to smart parking lots to identify a single license plate in a single lane, license plate recognition applied to intersection monitors must detect multiple license plates on multiple lanes. In addition, license plate recognition applied to intersection monitors faces many challenges, including too small license plates in the picture, unstable light sources, different shooting angles, blurred license plate characters in moving vehicles, and complex road conditions, advertising signs, traffic signs and road name indicator. To solve the above problems, this paper proposes a new two-stage methodology based on deep learning technology which first detects all the license plates in a picture and extracts the license plate images, and then performs character recognition on the license plate images using Convolutional Neural Networks. Through the two-stage approach, this method increases the proportion of characters in the picture, which in turn improves the character recognition accuracy. Experimental results show that the methodology achieves 98.23% of license plate detection rate and 97.38% of character recognition rate. The performance of the hierarchical methodology is about 25 fps. This methodology shows the superiority in both accuracy and performance in comparison with traditional license plate recognition systems.