ScholarWorks@숙명여자대학교

초록

As the reliability of autonomous driving systems becomes increasingly critical, research has been actively conducted on applying deep learning using GPUs as local edge devices. However, a major challenge remains: edge-based deep learning for autonomous driving often lacks the ability to quickly adapt to newly incoming data in real time. In this paper, we propose integrating Continual Learning (CL) algorithms into the deep learning training process on edge devices to enhance real-time adaptability and improve accuracy. Through experiments applying EWC and Rehearsal on an NVIDIA Jetson AGX-based edge environment, we demonstrate that CL algorithms can rapidly adapt to real-time environments without compromising accuracy.

키워드