자료유형  

 학위논문

Control Number  

0016931992

International Standard Book Number  

9798379651275

Dewey Decimal Classification Number  

900

Main Entry-Personal Name  

Cheng, Kelvin Bolin.

Publication, Distribution, etc. (Imprint  

[S.l.] : North Carolina State University., 2023

Publication, Distribution, etc. (Imprint  

Ann Arbor : ProQuest Dissertations & Theses, 2023

Physical Description  

1 online resource(82 p.)

General Note  

Source: Dissertations Abstracts International, Volume: 84-12, Section: A.

General Note  

Advisor: Simmons, Susan;Healey, Christopher G.

Dissertation Note  

Thesis (Ph.D.)--North Carolina State University, 2023.

Restrictions on Access Note  

This item must not be sold to any third party vendors.

Summary, Etc.  

요약The field of autonomous driving has grown tremendously in recent years thanks to the development of deep neural networks (DNNs) for computer vision tasks. As a cost-effective way of obtaining depth, DNN-based stereo matching plays a critical role in the development of self-driving systems. Although accuracy and processing speed have been the primary focus in recent stereo matching research, less emphasis has been placed on key aspects such as robustness, generalizability, and explainability, which are vital for the safety of autonomous vehicles. This thesis aims to address these shortcomings by exploring three crucial components of DNN-based stereo matching: 1) cost volume formation, 2) disparity distribution representation, and 3) intermediate layers in the cost aggregation stage.First, we devised a physically realizable adversarial attack technique that can easily manipulate the output of the current best-performing DNN-based stereo matching methods, presenting a significant danger to autonomous driving systems and exposing weaknesses in current DNN designs. To improve robustness against adversarial attacks, we combine a non-parametric cost volume with a parametric context feature map to replace the current parametric cost volume. Experiments confirm the effectiveness of our design in enhancing adversarial robustness and cross-domain generalizability.Second, we improve the single-view disparity distribution representation used in the current DNN-based stereo matching methods with a density-based volumetric representation. This allows for simultaneous generation of depth and occlusion maps for both views, as well as self-supervised learning without modifying DNN structures.Finally, to improve the interpretability of DNN-based stereo matching, we develop techniques to visualize the intermediate layers of the cost aggregation stage by converting intermediate tensors to disparity distributions. This offers a more intuitive understanding of how an initial cost volume is progressively transformed into a disparity distribution through the layers of a DNN. Experimentation using the visualization method reveals interesting insights and limitations in current DNNs by comparing different DNN-based stereo matching methods.

Subject Added Entry-Topical Term  

Maps.

Subject Added Entry-Topical Term  

Cameras.

Subject Added Entry-Topical Term  

Visualization.

Subject Added Entry-Topical Term  

Autonomous vehicles.

Subject Added Entry-Topical Term  

Neural networks.

Subject Added Entry-Topical Term  

Transportation.

Added Entry-Corporate Name  

North Carolina State University.

Host Item Entry  

Dissertations Abstracts International. 84-12A.

Host Item Entry  

Dissertation Abstract International

Electronic Location and Access  

로그인을 한후 보실 수 있는 자료입니다.

Control Number  

joongbu:643219