Electronic Theses and Dissertations

Author

Franz Parkins

Date

2020

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Electrical & Computer Engineering

Committee Chair

Eddie Jacobs

Committee Member

Deepak Venugopal

Committee Member

Aaron Robinson

Committee Member

Madhusudhanan Balasubramanian

Abstract

Artificial Intelligence has grown into an enormous field encompassing tasks formerly thought to require human intelligence. It encompasses the field of Machine Learning (ML) which uses data intensive learning algorithms to model complex data. Deep learning, using artificial neural networks, is the most prolific ML method. It can achieve responses to data that rival those of biological systems such as the visual cortexs response to motion. The field of computer vision has benefited from the availability of the deep learning methods throughout many common task such as object recognition, segmentation, and scene reconstruction but advancements in the area of Simultaneous Localization and Mapping (SLAM) has been gradual in comparison. The localization element of SLAM is computationally expensive and contains many years of refined domain knowledge that is not directly applicable to the neural approaches causing erroneous assumptions and error prone implementation. This is in part due to the temporal element of data associated with visual odometry. This research proposes a novel paradigm for feature engineering with respect to neural network architecture. It addresses some of the possible implementation assumptions. Internal feature engineering focuses on using a method that is internal to the neural architecture to expose the network to components of more favorable features. External feature engineering focuses on utilizing methods that operate outside of the neural architecture. The results show that these methods improve the task of iterative pose estimation for self-localization and are generally applicable for other tasks.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest

Share

COinS