In this paper, to secure the accuracy and real-time performance of the perception stage in autonomous driving, frame alignment was performed using a homography transformation technique. an efficient dataset was constructed based on normalized bounding box formats by applying the YOLOv5 architecture as an object recognition model. The trained PyTorch model was optimized for the EN675 NPU accelerator environment, enabling stable real-time detection and tracking of six major road objects(up to 32 objects per frame) across four channels of full HD camera inputs. This study is expected to provide high computational efficiency and reliability for edge devices in real-time autonomous driving applications.

Abstract
1. 서 론
2. 관련 연구
    2.1 딥러닝 기반 영상 이해
    2.2 자율주행 알고리즘
    2.3 자율주행 영상인식 및 사영변환 기술
3. 연구 방법 및 시스템 구현
    3.1 객체 인식 및 거리 측정 알고리즘 설계
    3.2 데이터셋 구축 및 어노테이션 포멧
    3.3 EN675 NPU 파이프라인 및 메모리 제어
    3.4 화면 레이아웃별 바운딩 박스 좌표 보정
    3.5 PyTorch-ONNX-NPU 모델 컴파일 과정
4. 결과 및 고찰
5. 결 론
References

• 유환신(Professor, Dept. of Automotive and Mechanical Eng, Howon University) | Yu Han Sin Corresponding author