EUVIMEDEUVIMED entdecken
Crossref

DESIGN AND IMPLEMENTATION OF AN END-TOEND DEEP LEARNING SYSTEM FOR AUTONOMOUS DRIVING: FROM VISUAL PERCEPTION TO INTELLIGENT VEHICLE CONTROL

Sungho Kim, Khairul Anam, Md Nafis Azad Nobel, Pritom Das, Mohammad Somon Sikder, Hemayet Uddin Himel, Yearanoor Khan, Seohyun Yang, Saeromi Kim, Ibtisum Ahmed Nihal

International Journal of Grid Computing & Applications · 28.03.2026 · 17 (1) · 93-99

DOI: 10.5121/ijgca.2026.17108 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Grid Computing & Applications, online erschienen am 28.03.2026. Verfasst von Sungho Kim und 9 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

End-to-end deep learning has become an important and significant driver of autonomous driving study because it substitutes an intricate multi-stage sense; plan; act pipeline with a single extremely multi-faceted differentiable model that can be trained end-to-end from sensorial input all the way to control outcome. Though end-to-end learning is profitable and useful relative to classical modular pipelines in that it can learn to compress the visual information in an end-to-end fashion and consequently learn task-specific latent representations optimized for driving, the identical design decisions that favor holistic optimization furthermore complicate analysis and provide difficulty on the way toward safe deployment: distribution shift in imitation learning, interpretability, debugging, and safety in closed-loop control. This document describes a real-world end-to-end CNN forsteering angle prediction trained by means of supervised imitation learning on the Udacity Behavioral Cloning dataset recorded in the Udacity simulator . Our model is largely inspired by the NVIDIA end-to-end architecture (frequently called PilotNet), which performs CNN-to-steering prediction employing sole camera image input. We employ a perceptual-tocontrol training system that uses targeted region-of-interest cropping, input normalization, and data augmentation (brightness modifications, horizontal flipping, and spatial shifting) to generalize better from variation in lighting and camera standpoint. Our proposed network on the held-out test set results MSE=0.0166 and MAE=0.0682, which is quiterobust open-loop prediction accuracy for steering regression given the running conditions of the dataset. We will moreover elaborate on the reasons that robust open-loop regression metrics do not fully characterize closed-loop driving safety, and the safety and interpretability instruments (for example, causal attention maps and safety monitors) which we describe that can supplement end-to end control policies in actual deployment. Index Terms: Autonomous Driving, End-to-End Learning, Imitation Learning, Behavioral Cloning, Convolutional Neural Networks, Direct Perception, Explainable AI, AI Safety

Zitieren

APA

Kim, S., Anam, K., Nobel, M. N. A., Das, P., Sikder, M. S., Himel, H. U., Khan, Y., Yang, S., Kim, S., & Nihal, I. A. (2026). DESIGN AND IMPLEMENTATION OF AN END-TOEND DEEP LEARNING SYSTEM FOR AUTONOMOUS DRIVING: FROM VISUAL PERCEPTION TO INTELLIGENT VEHICLE CONTROL. International Journal of Grid Computing & Applications, 17(1), 93-99. https://doi.org/10.5121/ijgca.2026.17108

Vancouver

Kim S, Anam K, Nobel MNA, Das P, Sikder MS, Himel HU, et al. DESIGN AND IMPLEMENTATION OF AN END-TOEND DEEP LEARNING SYSTEM FOR AUTONOMOUS DRIVING: FROM VISUAL PERCEPTION TO INTELLIGENT VEHICLE CONTROL. International Journal of Grid Computing & Applications. 2026;17(1):93-99. doi:10.5121/ijgca.2026.17108

BibTeX
@article{Kim2026,
  author = {Kim, Sungho and Anam, Khairul and Nobel, Md Nafis Azad and Das, Pritom and Sikder, Mohammad Somon and Himel, Hemayet Uddin and Khan, Yearanoor and Yang, Seohyun and Kim, Saeromi and Nihal, Ibtisum Ahmed},
  title = {DESIGN AND IMPLEMENTATION OF AN END-TOEND DEEP LEARNING SYSTEM FOR AUTONOMOUS DRIVING: FROM VISUAL PERCEPTION TO INTELLIGENT VEHICLE CONTROL},
  journal = {International Journal of Grid Computing & Applications},
  year = {2026},
  volume = {17},
  number = {1},
  pages = {93--99},
  doi = {10.5121/ijgca.2026.17108}
}
Kostenlos starten →