Discipline: Computer Sciences and Information Management
Subcategory: Computer Science & Information Systems
Session: 1
Room: Private Dining
Grace Vincent - North Carolina State University
Co-Author(s): Gabriel Ohmes, Fayetteville State University, Fayetteville, NC; Nathan Couch, Fayetteville State University, Fayetteville, NC; Ethan Patten, North Carolina State University, Raleigh, NC; Devon Kennedy, Cross Creek Early College, Fayetteville, NC
Autonomous systems require robots, agents, that can understand and interpret their surroundings. The goal is not only for an agent to be able to detect whether an object is present, but for it to have the ability to comprehend what that object is. This will allow them to enhance their decision-making processes and potentially minimize risk in downstream tasks. Such skills can be utilized in a variety of fields such as search and rescue and exploration. As developments in an individual agent’s ability to perceive its surroundings progress, we can extend to using multiple agents that work together in a system. This multi-agent system (MAS) enables multiple robots to collaborate to explore the same environment efficiently for full scene understanding and analysis. The deployment of a MAS also allows for the capture of additional information and the potential for exploration of larger environments. Collaborative MASs require correspondence identification that enables each agent to refer to the same objects (i.e., hazards, science targets, etc.) within their own field of view. Yet, this process of correspondence identification has challenges that arise when agents experience non-covisibility of objects, i.e., objects that are not visible in all fields of view. Using state-of-the-art stereovision imaging systems, the ZED 2i cameras, we have captured RGB+D information across multiple agents to develop our correspondence identification algorithm. We initially attempted to implement an assignment algorithm, Hungarian algorithm, that utilized a Hausdorff distance. This assignment algorithm had difficulty handling non-covisible objects and was generally uncertain about the correspondence when the number of agents and objects increased. Thus, we have pivoted to developing a graph matching algorithm to identify the similarities between objects in different fields of view. This graph matching algorithm will better handle the challenges of correspondence identification, non-covisibility, with higher certainty than previously attempted assignment algorithms because of the ability to interpret information of an object’s attributes and its position in space relative to its own and other agents’ frames. Additionally, compared to the assignment algorithm, the graph matching approach will produce less computationally expensive results, which can be critical for resource-limited agents or tasks. This enhanced correspondence identification develops a world reference frame so that the agents within the system can perform their downstream tasks strategically and efficiently.
Funder Acknowledgement(s): This work was supported in part by a NASA MUREP Space Technology Artemis Research (M-STAR) Implementation Grant.
Faculty Advisor: Sambit Bhattacharya, sbhattac@uncfsu.edu
Role: In this research team I was acting as the lead student researcher. I oversaw tasking the team to review required materials to develop a base knowledge of the problem and different methodologies for solving these issues. Also, I ensured that our team continually made progress and met deadlines for the project. As a team member, I worked on the development of the different algorithms and primary developer for the graph matching solution. Additionally, I aided my team members in the setup of the environment and utilities necessary for conducting the experiments and data capture.