Recent advances in the field of deep learning technologies have made it possible to develop practical video analysis systems with embedded platform that are more accurate and faster than prior embedded systems based on pattern analysis technology. In video analysis applications, object detection, face recognition, action recognition and super resolution technology is the most important functions. In this paper, we show action recognition and super resolution on embedded deep learning system. It is discovered that ResNet34 structure with 16 frames analysis is most profitable for speed and accuracy and Attention based super resolution method is enough for real-time processing. Both deep learning models optimized for speed and accuracy are operated on embedded system with Intel NPU at real-time. We introduce main technologies in chapter 1. Proposed method is shown in chapter 2 and the result is shown in chapter 3.