請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/93109
標題: | 夜間暨低光源下自駕即時影像分割、深度及優化模組 Low Light Enhancement, Depth and Image Segmentation in Real-time and Nighttime Driving Scene Video for Autonomous Vehicles |
作者: | 盧怡臻 I-Chen Lu |
指導教授: | 貝蘇章 Soo-Chang Pei |
關鍵字: | 影像,影片,夜間,即時,深度學習,機器學習,輕量化,低光源亮度增強,領域適應,深度估計,光流,語意分割,實例分割,全景分割,全景深度分割, Image,Video,Nighttime,Real-time,Deep Learning,Machine Learning,Lightweight,Low-Light Enhancement,Domain Adaption,Depth Estimation,Optical Flow,Semantic Segmentation,Instance Segmentation,Panoptic Segmentation,Panoptic Depth Segmentation, |
出版年 : | 2024 |
學位: | 碩士 |
摘要: | 隨著自動駕駛技術的進步,夜間駕駛場景的挑戰變得越來越明顯。在低光源條件下,影像清晰度和可見度受到嚴重影響,對自動駕駛系統的性能和安全性構成挑戰。本研究探索低光增強、深度估計和影像分割等技術,以提高自動駕駛汽車在夜間駕駛場景中的表現。首先,採用低光源增強技術增強夜間駕駛場景影像的亮度和對比度,進而改善駕駛者和系統對環境的感知。其次,深度估計技術可以準確估計夜間場景中物體的距離和位置信息,為自動駕駛系統提供關鍵的環境感知數據。最後,影像分割技術可以精確識別和分割夜間駕駛場景中的各種物體,為自動駕駛汽車提供更準確、更可靠的環境理解。除此之外,我們提出了名為 Panoptic-LMFFNet、NLP-Net、Panoptic-DLMFFNet 和 NLPD 的模型,它們分別處理白天和夜間的即時全景分割和全景深度分割。這項研究的結果表明,結合微光增強、深度估計和影像分割技術可以顯著提高自動駕駛汽車在夜間駕駛場景中的性能和安全性。 As autonomous vehicle technology advances, the challenges of nighttime driving scenes become increasingly apparent. In low-light conditions, image clarity and visibility are severely compromised, posing challenges to the performance and safety of autonomous driving systems. This study explores techniques such as low-light enhancement, depth estimation, and image segmentation to improve the performance of autonomous vehicles in nighttime driving scenes. Firstly, low-light enhancement techniques are employed to enhance the brightness and contrast of images in nighttime driving scenes, thereby improving the perception of the environment by both drivers and systems. Secondly, depth estimation techniques accurately estimate the distance and location information of objects in nighttime scenes, providing crucial environmental perception data for autonomous driving systems. Finally, image segmentation techniques precisely identify and segment various objects in nighttime driving scenes, providing a more accurate and reliable environmental understanding of autonomous vehicles. Otherwise, we proposed models named Panoptic-LMFFNet, NLP-Net, Panoptic-DLMFFNet and NLPD, and they deal with real-time panoptic segmentation and panoptic depth segmentation in daytime and nighttime respectively. The results of this study demonstrate that combining low-light enhancement, depth estimation, and image segmentation techniques significantly improves the performance and safety of autonomous vehicles in nighttime driving scenes. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/93109 |
DOI: | 10.6342/NTU202401061 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-2.pdf 此日期後於網路公開 2024-09-01 | 78.2 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。