Technology trend 4.0 has created a revolution in devices integrating artificial intelligence (AI) to solve everyday problems, including license plate recognition technology. Currently, many organizations in the world and in Vietnam are developing this technology. However, usage is limited due to different requirements and specificities, such as requiring high computing power, high fees, high latency, and unstable connection. Since the characters and formats used in license plates vary greatly from country to country, it is still challenging to develop an automatic license plate number recognition system. This work aims to develop a license plate recognition system using embedded devices that is suitable for Vietnamese license plates. Our method focus on ensuring stable performance on devices with low processing power and stability in real time. We propose to use a two-step license plate recognition system, the first step uses the Single shot multibox detector algorithm based on the MobileNetV2 as backbone to detect license plates. Next, YOLOv8-nano is used to recognize characters on license plates, trained on the Vietnamese license plate data set, using Raspberry Pi 4 hardware and Raspberry Pi Camera Module V2 8MP. The system achieves an average recognition accuracy of 95.68% and an average stable execution time of 0.478 s. The system has successfully provided a cost-effective, scalable and widely applicable solution in Vietnam license plate environment.