TRANSFER LEARNING OF A NEURAL NETWORK ONBOARD AN UNMANNED SURFACE VEHICLE USING SYNTHETIC DATA
Loading...
Authors
Fullilove, Trey M.
Subjects
USV
neural network
object detection
synthetic data
neural network
object detection
synthetic data
Advisors
Kragelund, Sean P.
Date of Issue
2024-12
Date
Publisher
Monterey, CA; Naval Postgraduate School
Language
Abstract
This thesis explores the use of transfer learning and synthetic data to enhance the performance of a convolutional neural network (CNN) for maritime vessel detection, focusing on autonomous unmanned surface vehicles (USVs). The challenge addressed is the limited availability of labeled nighttime data, which is crucial for detecting vessels in low-visibility conditions. The model was initially trained on real-world data to detect the Mokai USV and other vessels. Synthetic daytime data was then used for transfer learning to diversify the dataset with more vessel types and conditions. Additionally, histogram equalization was applied to images from the Sionyx NightWave camera to adapt the model for nighttime detection without requiring extensive real-world labeled nighttime data. The model’s performance was evaluated using average precision (AP), recall, and miss rate metrics across real-world, synthetic daytime, and synthetic nighttime datasets. While the model performed well under daytime conditions, its performance dropped for nighttime detection, particularly for smaller objects. These results highlight the challenge of domain adaptation, but also demonstrate the potential of synthetic data and transfer learning for addressing the scarcity of labeled data in maritime environments. This approach offers a cost-effective solution for improving autonomous USV operations in diverse conditions.
Type
Thesis
Description
Series/Report No
Organization
Identifiers
NPS Report Number
Sponsors
Funder
Format
Citation
Distribution Statement
Distribution Statement A. Approved for public release: Distribution is unlimited.
Rights
This publication is a work of the U.S. Government as defined in Title 17, United States Code, Section 101. Copyright protection is not available for this work in the United States.