Journal of the Korea Society of Computer and Information (한국컴퓨터정보학회논문지)

Korean Society of Computer Information (한국컴퓨터정보학회)
권호 표지
DOI QR코드

DOI QR Code

A Study on the Impact of Noise on YOLO-based Object Detection in Autonomous Driving Environments

  • Ra Yeong Kim (Dept. of Cyber Security, Pai Chai University) ;
  • Hyun-Jong Cha (Dept. of Software Engineering, Pai Chai University) ;
  • Ah Reum Kang (Dept. of Cyber Security, Pai Chai University)
  • Received : 2024.08.01
  • Accepted : 2024.09.25
  • Published : 2024.10.31
Download PDF
⟨ Previous Next ⟩

Abstract

Noise caused by adverse weather conditions in data collected during autonomous driving can lead to object recognition errors, potentially resulting in critical accidents. While this risk is widely acknowledged, there is a lack of research that quantitatively and systematically analyzes it. Therefore, this study aims to examine and quantify the extent to which noise affects object detection in autonomous driving environments. To this end, we utilized the YOLO v5 model trained on unprocessed datasets. The test data were divided into noise ratios of 0% (Original), 20%, 40%, 60%, and 80%, and the detection results were evaluated by constructing a Confusion Matrix. Experimental results show that as the noise ratio increases, the True Positive (TP) rate decreases, and the F1-score also significantly drops across all noise levels, specifically from 0.69 to 0.47, 0.29, 0.18, and 0.14. These findings are expected to contribute to enhancing the stability of autonomous driving technology. Future research will focus on collecting real datasets that include naturally occurring noise and developing more effective noise removal techniques.

자동차의 자율주행 시 수집되는 데이터에 악천후 등의 원인에서 발생하는 노이즈는 객체 인식의 오류를 초래하여 자율주행에 각종 사고 등 치명적인 결과를 낳을 가능성이 크다. 이러한 위험성은 널리 인식되고 있지만, 이를 정량화하고 체계적으로 분석한 연구는 부족한 실정이다. 따라서 본 연구에서는 자율주행 환경에서 노이즈가 적용되는 비율에 따라 객체를 인식하는 것에 어느 정도의 영향을 미치는지 살펴보고 수치화하고자 한다. 이를 위해 가공을 하지 않은 데이터셋으로 학습한 YOLO v5 모델을 사용한다. 이를 통해 노이즈 비율을 0%(Original), 20%, 40%, 60%, 80%로 나눈 Test Data를 detection 하여 결과를 확인하고 Confusion Matrix를 구성하여 실험 결과를 평가했다. 실험 결과, 노이즈의 비율이 높아질수록 TP(True Positive)의 비율이 낮아지는 것을 확인할 수 있었으며, F1-score 또한 0.69, 0.47, 0.29, 0.18, 0.14로 적은 노이즈 비율뿐 아니라 모든 비율에서 수치가 큰 폭으로 떨어진다는 것을 알 수 있다. 이를 통해 자율주행 기술의 안정성을 높일 수 있는 기반이 될 것이라 기대되며 향후에는 자연적으로 발생한 노이즈가 포함된 실제 데이터셋을 수집하고 발생된 노이즈를 더욱 효과적으로 제거할 수 있는 연구를 진행할 예정이다.

Keywords

Acknowledgement

This research was supported by the MSIT(Ministry of Science and ICT), Korea, under the Innovative Human Resource Development for Local Intellectualization support program(IITP-2024-RS-2022-00156334) supervised by the IITP(Institute for Information & communications Technology Planning & Evaluation)

References

  1. Im. Y. Jung, "A Study on The Dangers and Their Countermeasures of Autonomous Vehicle," Vol. 20, no. 6, pp. 90-98, Jan. 2020, DOI: 10.5392/JKCA.2020.20.06.090.
  2. Soonhong Kwon, and Jong-Hyouk Lee, "Self-driving Car Security Threats and Technology Trends," REVIEW OF KIISC, Vol. 30, No. 2, pp. 31-39, 2020.
  3. Joseph Redmon, Santosh Divvala, Ross Girshick and Ali Farhadi, "You Only Look Once:Unified, Real-Time Object Detection," arXiv:1506.02640, May 2016, DOI: 10.48550/arXiv.1506.02640
  4. Yan, Bin, Pan Fan, Xiaoyan Lei, Zhijie Liu, and Fuzeng Yang, "A Real-Time Apple Targets Detection Method for Picking Robot Based on Improved YOLOv5," Remote Sensing 13, no. 9: 1619, Apr. 2021, DOI: 10.3390/rs13091619
  5. Joon-Yong Kim, "A Comparative Study on the Characteristics of Each Version of Object Detection Model YOLO," Proceedings of the Korean Society of Computer Information Conference, Vol. 31, No. 2, pp.75-78, Jeju, Korea, Jul. 2023.
  6. Jaewon Ahn, DaeJin Kim, and Jun Kyun Choi, "A Study on Modeling the Object Detection Accuracy According to the Image Resolution," Proceedings of Symposium of the Korean Institute of communications and Information Sciences, pp.64-65, Youngpyoung, Korea, Feb. 2020.
  7. Sungick Kong, Sang-Seol Lee, Sung-Joon Jang, and Byeongho Choi, "Korean Traffic-Sign Detection and Recognition using YOLO," Autumn Annual Conference of IEIE, pp.832-0834, Incheon, Korea, Aug. 2018.
  8. Jisoo Tak, and Sang-Woong Lee, "Noise-Augmented Object Detection: Improving Object Detection Accuracy in Noisy Image," Korean Institute of Next Generation Computing, pp.141-144, Kyungnam University, Korea, Jun. 2023.
  9. Junnan Li, Caiming Xiong, Richard Socher, and Steven Hoi, "Towards Noise-resistant Object Detection with Noisy Annotations," arXiv preprint arXiv2003.01285, Mar. 2020. DOI: 10.48550/arXiv.2003.01285
  10. Hwe-soo Park, Pil-joong Kim, and Seong-Joo Lee, "Change of Recognition Rate in YOLO v4 According to Image Quality," Summer Annual Conference of IEIE, pp.1984-1986, Jeju, Korea, Aug. 2020.
  11. EunJin Jeon, YoonSeok Cha, Seongyeol An, and ByungRae Cha, "Comparison and Analysis of YOLO Object Detection According to Image Quality Clarity," 2021 KICS Fall Conference, pp.513-515, Yeosu, Korea, Nov. 2021.
  12. Nguten Thi My Xuyen, "Object Detection and Speed Prediction in Night Environment Using Deep Learning", Master's thesis, National Pukyong National University, Pusan, Feb. 2024.
  13. Na-Hyoun Kim, Donghoon Lee, Junwoo Kim, Kee-Young Kwahk, and Namgyu Kim, "Multiple Denoising Scheme for Effective Nighttime Object Detection", Proceedings of KIIT Conference, pp.88-90, Jeju, Korea, Nov. 2021.
  14. Jieun Kim, Chan-young Jung, Woong Hwang, Don June Lim, and Hyeong-ju Noh, "YOLO based Obect detection for Autonomous driving and Collison Warning", proceedings of HCI Korea 2023, pp. 1,107-1,110, Gangwon, Korea, Feb. 2023.
  15. Kana Kim, and Hakil Kim, "AI Model-Based Automated Data Cleaning for Reliable Autonomous Driving Image Datasets", Journal of Broadcast Engineering, Vol. 28, No. 3, pp.302-313, May. 2023.
  16. Choi, Yoonjo, Lee, Gyeong Gwan, & Hong, Seunghwan, "A study on Automatic Algorithm of 3D Dynamic Object Annotation for Autonomous Driving AI Training Data Construction", KSCE 2023 convention, pp.672-673, Oct. 2023.
  17. Seongho Son, and Changsun Ahn, "Denoising LiDAR Point Cloud Using Higher Order Singular Value Decomposition in Adverse Weather Condition", KSAE 2021 Annual Spring Conference, pp. 435-435, Pyeongchang, Korea, Jun. 2021.
  18. Taresh Sarvest Sharan, Romel Bahttacharjee, Shiru Sharma, and Neeraj Sharma, "Evaluation of Deep Learning Methods (DnCNN and U-Net) for Denoising of Heart Auscultation Signals", 2020 3rd International Conference on Communication System, Computing and IT Applications (CSCITA), pp. 151-155, Mumbai, India, Jul 2020. DOI:10.1109/CSCITA47329.2020.9137813.