메뉴 건너뛰기
소속 기관 / 학교 인증
인증하면 논문, 학술자료 등을  무료로 열람할 수 있어요.
한국대학교, 누리자동차, 시립도서관 등 나의 기관을 확인해보세요
(국내 대학 90% 이상 구독 중)
고객센터 ENG
주제분류

논문 기본 정보

저자정보
(서울과학기술대학교) (서울과학기술대학교) (서울과학기술대학교)
저널정보
제어로봇시스템학회 제어로봇시스템학회 논문지 제어로봇시스템학회 논문지 제31권 제4호
오류 신고하기

피인용 0

검색

    초록·키워드

    Recently developed autonomous driving systems based on deep learning typically operate through modular architectures, where separate modules perform distinct individual tasks. While the UniAD framework proposed in the “Planning-oriented Autonomous Driving” paper addresses the limitations of modular approaches through a unified architecture, its complex transformer structure requires substantial computational resources to function. This paper proposes a lightweight version of UniAD to improve the accessibility of multimodal learning. We reduce the computational complexity by lowering the number of transformer layers and queries, the dimensions, and the BEV spatial resolution. Additionally, we optimize memory usage by limiting sampling queries and enabling page-locked memory settings. Experiments with two versions of the lightweight architecture show significant memory reductions: up to 79.92% in Stage 1 and 38.81% in Stage 2 compared with the original UniAD architecture (52.3 GB and 16.67 GB, respectively). Although the lightweight model suffers an overall performance degradation, we discover that progressive resolution expansion during training can enhance its feature extraction capability, particularly in the initial low-resolution learning phase.

    최근 본 자료 전체보기

      UCI(KEPA) : I410-151-25-02-092782145