ScholarWorks@성균관대학교

초록

With the recent active research on autonomous driving technology, the importance of deep learning models that accurately recognize the surrounding environment in real time has been highlighted. In particular, it is essential to select a model that considers not only accuracy but also real-time and hardware efficiency in order to mount it on hardware with limited computational resources. In this paper, we aim to select the lightweight segmentation model that is most suitable for RTL (Register - Transistor Level) implementation for the construction of lane recognition system of autonomous driving system. To this end, we compared-analyzed the representative real-time models Fast-SCNN, BiSeNetV2, and DDRNet in the same learning environment. As a result of the experiment, Fast-SCNN showed the best performance among them and showed a advantage in terms of hardware efficiency. Therefore, this study finally selected Fast-SCNN as the semi-segmentation model of the next-generation autonomous driving system.

키워드