Dissertations, Theses, and Capstone Projects

Date of Degree

2-2018

Document Type

Dissertation

Degree Name

Ph.D.

Program

Computer Science

Advisor

Scott Dexter

Committee Members

Robert Haralick

Ioannis Stamos

Suzanne McIntosh

Subject Categories

Robotics | Signal Processing | Systems and Communications | Systems Engineering

Keywords

Computer Vision, Localization, Place Recognition, Feature Extraction, Feature Description

Abstract

This dissertation makes contributions to the problem of Long-Term Appearance-Based Place Recognition. We present a framework for place recognition in a collaborative scheme and a method to reduce the impact of dynamic objects on place representations. We demonstrate our findings using a state-of-the-art place recognition approach.

We begin in Part I by describing the general problem of place recognition and its importance in applications where accurate localization is crucial. We discuss feature detection and description and also explain the functioning of several place recognition frameworks.

In Part II, we present a novel framework for collaboration between agents from a pure appearance-based place recognition perspective. Using this framework, multiple agents can efficiently share partial or complete knowledge about places and benefit from their teamwork. This collaborative framework allows agents with limited storage and memory capacity to become useful in environment exploration tasks (for instance, by enabling remote recognition); includes procedures to manage an agent’s memory load and distributes knowledge of places across agents; allows the reuse of knowledge from one agent to another; and increases the tolerance for failure of individual agents. Part II also defines metrics which allow us to measure the performance of a system that uses the collaborative framework.

Finally, in Part III, we present an innovative method to improve the recognition of places in environments densely populated by dynamic objects. We demonstrate that we can improve the recognition performance in these environments by incorporating high- level information from dynamic objects. Tests conducted using a synthetic dataset show the benefits of our approach. The proposed method allows the system to significantly improve the recognition performance in the photo-realistic dataset while reducing storage requirements, resulting in up to 23.7 percent less storage space than the state-of-the-art approach that we have extended; smaller representations also reduced the time required to match places. In Part III, we also formulate the concept of a valid place representation and determine the quality of the observation based on dynamic objects present in the agent’s view.

Of course, recognition systems that are sensitive to dynamic objects incur additional computational costs to recognize those objects. We show that this additional cost is outweighed by the benefits that incorporating dynamic object detection in the place recognition pipeline. Our findings can be used in many applications, including applications for navigation, e.g. assisting visually impaired individuals with navigating indoors, or autonomous vehicles.

Share

COinS