자료유형  

 학위논문

Control Number  

0017162943

International Standard Book Number  

9798384345046

Dewey Decimal Classification Number  

001

Main Entry-Personal Name  

Zhang, Miao.

Publication, Distribution, etc. (Imprint  

[S.l.] : Stanford University., 2024

Publication, Distribution, etc. (Imprint  

Ann Arbor : ProQuest Dissertations & Theses, 2024

Physical Description  

141 p.

General Note  

Source: Dissertations Abstracts International, Volume: 86-04, Section: B.

General Note  

Advisor: Rock, Stephen.

Dissertation Note  

Thesis (Ph.D.)--Stanford University, 2024.

Summary, Etc.  

요약This thesis introduces T-STORE (a long-term tracker with dynamic DCF Template STORE(-age) for target recovery), an algorithm designed to enable long-duration vision-based tracking of highly deformable ocean midwater animals using autonomous underwater vehicles.Current midwater tracking practices typically employ stereo blob tracking algorithms to accomplish this task. These systems have proven to be highly effective when the target animal remains within the camera's field-of-view, achieving tracking durations of multiple hours. These systems fail, however, when disruptions occur, such as the target temporarily moving out of view or being occluded by other objects. As a result, they are unsuited for applications requiring tracking durations greater than 24 hours. Addressing the challenge of target re-identification post-disruptions is essential for extending tracking durations.T-STORE presents a solution to the target re-identification problem by fusing the stereo blob tracking system with a visual template-based learning system built on a pool of online-acquired Discriminative Correlation Filer (DCF) templates. Instead of employing an offline-trained Convolutional Neural Network (CNN) detector, T-STORE utilizes an online template-based learning approach to address the scarcity of large-scale data on midwater targets and to enable the tracking of previously unseen targets-of-opportunity.T-STORE is the first to integrate stereo blob tracking with Discriminative Correlation Filters (DCFs). It introduces a novel target re-identification pipeline, leveraging a unique template matching metric and exploiting target information collected during online learning. Additionally, T-STORE introduces a set of lightweight deep features optimized for DCF-based template matching, ensuring adaptability to appearance changes and reducing the required number of templates, which in turn enhances suitability for real-time applications.The performance of T-STORE is demonstrated using field data containing challenging tracking conditions that lead to failures in the stereo blob tracking algorithm. The experiments include a comparison with FuCoLoT (a Fully Correlational Long-Term Tracker), an accepted leader in DCF template-based tracking algorithms. T-STORE outperformed FuCoLoT in both standard tracking metrics and target recovery performance. Specifically, T-STORE exhibited a target recovery success rate exceeding 80%, double that of FuCoLoT, along with an overall F-score of over 0.8, indicating its promising potential for achieving the long duration tracking goal.

Subject Added Entry-Topical Term  

Visualization.

Subject Added Entry-Topical Term  

Decision making.

Subject Added Entry-Topical Term  

Autonomous underwater vehicles.

Subject Added Entry-Topical Term  

Naval engineering.

Subject Added Entry-Topical Term  

Aerospace engineering.

Added Entry-Corporate Name  

Stanford University.

Host Item Entry  

Dissertations Abstracts International. 86-04B.

Electronic Location and Access  

로그인을 한후 보실 수 있는 자료입니다.

Control Number  

joongbu:657424