Access to master’s theses is restricted to current ERAU students, faculty, and staff.
Date of Award
Thesis - Open Access
Master of Science in Mechanical Engineering
Dr. Charles Reinholtz
Dr. Sathya Gangadharan
Research into object classification has led to the creation of hundreds of databases for use as training sets in object classification algorithms. Datasets made up of thousands of cars, people, boats, faces and everyday objects exist for general classification techniques. However, no commercially available database exists for use with detailed classification and categorization of marine vessels commonly found in littoral environments. This research seeks to fill this void and is the combination of a multi-stage research endeavor designed to provide the missing marine vessel ontology. The first of the two stages performed to date introduces a novel training database called the Lister Littoral Database 900 (LLD-900) made up of over 900 high-quality images. These images consist of high-resolution color photos of marine vessels in working, active conditions taken directly from the field and edited for best possible use. Segmentation masks of each boat have been developed to separate the image into foreground and background sections. Segmentation masks that include boat wakes as part of the foreground section are the final image type included. These are included to allow for wake affordance detection algorithms rely on the small changes found in wakes made by different moving vessels. Each of these three types of images are split into their respective general classification folders, which consist of a differing number of boat categories dependent on the research stage.
In the first stage of research, the initial database is tested using a simple, readily available classification algorithm known as the Nearest Neighbor Classifier. The accuracy of the database as a training set is tested and recorded and potential improvements are documented. The second stage incorporates these identified improvements and reconfigures the database before retesting the modifications using the same Nearest Neighbor Classifier along with two new methods known as the K-Nearest Neighbor Classifier and the Min-Mean Distance Classifier. These additional algorithms are also readily available and offer basic classification testing using different classification techniques. Improvements in accuracy are calculated and recorded. Finally, further improvements for a possible third iteration are discussed.
The goal of this research is to establish the basis for a training database to be used with classification algorithms to increase the security of ports, harbors, shipping channels and bays. The purpose of the database is to train existing and newly created algorithms to properly identify and classify all boats found in littoral areas so that anomalous behavior detection techniques can be applied to determine when a threat is present. This research represents the completion of the initial steps in accomplishing this goal delivering a novel framework for use with littoral area marine vessel classification. The completed work is divided and presented in two separate papers written specifically for submission to and publication at appropriate conferences. When fully integrated with computer vision techniques, the database methodology and ideas presented in this thesis research will help to provide a vital new level of security in the littoral areas around the world.
Scholarly Commons Citation
Lister, Robert Andrew, "Classification of Marine Vessels in a Littoral Environment Using a Novel Training Database" (2011). Theses - Daytona Beach. 122.