請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88098
標題: | 駕駛場景中的影像辨識:三維物件辨識與影像分割 Image Recognition in Driving Scene: Monocular 3D Object Detection and Image Segmentation |
作者: | 游家權 Jia-Quan Yu |
指導教授: | 貝蘇章 Soo-Chang Pei |
關鍵字: | 全景分割,物件辨識,安全性指標,駕駛場景,深度估計,三維物件辨識,語意分割,實例分割, Panoptic Segmentation,Object Detection,Safety Metric,Driving Scene,Depth Estimation,3D Object Detection,Semantic Segmentation,Instance Segmentation, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 影像辨識演算法在自駕車領域中扮演著關鍵的角色,它使得計算機能夠感知周圍環境並確保駕駛安全。本研究聚焦於自動駕駛場景中的物件檢測和影像分割任務的相關深度學習技術,針對三維物件辨識,本研究提出基於場景的數據擴增方法、錨點採樣技巧以及能夠改進特徵提取能力的相機視角卷積模組,本研究於KITTI3D數據集的測試中取得了顯著的23.9%平均準確度。此外,針對影像分割任務,我們提出了整合深度估計和全景分割的新任務並開發了Panoptic-DepthLab網絡,以實現分割和深度估計任務的共同訓練。本研究旨在提供自動駕駛領域中的影像識別演算法可能的改進方向,並設法提高物件檢測和影像分割的準確性和效率。 Image recognition plays a crucial role in autonomous driving, enabling computers to perceive the surrounding environment and ensure driving safety. In this thesis, we focus on both detection and segmentation tasks in driving scenarios, leveraging deep-learning techniques. Our work introduces novel approaches to enhance the performance of 3D object detection, including a scene-aware data augmentation method, a depth-aware anchor sampling technique, and a perspective-aware convolutional module that improves feature extraction capabilities. By combining these methods, we achieve a significant AP of 23.9% on the easy benchmark of the KITTI3D dataset. Furthermore, we propose a novel integration of depth estimation and panoptic segmentation on a color map. We also introduce the Panoptic-DepthLab network, which enables joint training of segmentation and depth estimation tasks. Our research aims to advance the field of image recognition in autonomous driving, improving the accuracy and efficiency of object detection and segmentation algorithms. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88098 |
DOI: | 10.6342/NTU202301439 |
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-111-2.pdf | 100.52 MB | Adobe PDF | 檢視/開啟 |
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。