Early Classification of Intent for Maritime Domains Using Multinomial Hidden Markov Models

Logan Carlson; Dalton Navalta; Monica Nicolescu; Mircea Nicolescu; Gail Woodward

doi:10.3389/frai.2021.702153

Early Classification of Intent for Maritime Domains Using Multinomial Hidden Markov Models

Front Artif Intell. 2021 Oct 7:4:702153. doi: 10.3389/frai.2021.702153. eCollection 2021.

Authors

Logan Carlson¹, Dalton Navalta¹, Monica Nicolescu¹, Mircea Nicolescu¹, Gail Woodward²

Affiliations

¹ Department of Computer Science and Engineering, University of Nevada Reno, Reno, NV, United States.
² NASA Jet Propulsion Laboratory, Pasadena, CA, United StAtes.

Abstract

The need for increased maritime security has prompted research focus on intent recognition solutions for the naval domain. We consider the problem of early classification of the hostile behavior of agents in a dynamic maritime domain and propose our solution using multinomial hidden Markov models (HMMs). Our contribution stems from a novel encoding of observable symbols as the rate of change (instead of static values) for parameters relevant to the task, which enables the early classification of hostile behaviors, well before the behavior has been finalized. We discuss our implementation of a one-versus-all intent classifier using multinomial HMMs and present the performance of our system for three types of hostile behaviors (ram, herd, block) and a benign behavior.

Keywords: behavior recognition; intent recognition; machine learning; maritime; multinomial HMM.