YIP Active Perception and Learning for Adaptive Underwater Surveying and Automatic Target Recognition
Abstract
The goal of this proposed project is to advance automatic target recognition (ATR) capabilities to enable automated target detection, classification, and reconstruction in complex marine environments. The key innovation will be to leverage recent advances in implicit neural representations and uncertainty quantification for deep learning to develop a novel paradigm for adaptive ATR surveying.Implicit neural representations have demonstrated impressive capabilities to enable novel view synthesis and dense scene reconstruction, which can improve downstream robot perception tasks. We will extend this prior work to enable learning implicit neural representations from side scan sonar sensors. We will also address critical challenges to deploying implicit neural representations in realfield environments. To achieve these goals, this project proposes the following research aims: (i) to develop novel machine learning algorithms to enable automated target detection and uncertainty estimationfrom side scan sonar imagery, (ii) to leverage target predictions and associated uncertainty to learn implicit neural representations for side scan sonar imagery from limited input views to enable novel view synthesis and 3D reconstruction of the target, and (iii) to develop an uncertainty-aware active perception framework that leverages implicit neural representations for next-best view selection to acquire additional informative views of the target. The developed methods will be evaluated across simulated datasets, lab experiments, and real underwater environments to provide qualitative and quantitative evaluation of performance across varying environmental conditions. The expected outcome is to enable adaptive, autonomous surveying across the full mission pipeline including target detection and online replanning to capture additionalinformative views, while minimizing additional survey time. This will be transformative technology that will enable efficient and effective ATR in challenging operational scenarios, with relevance to future naval systems with applications in mine countermeasures,anomaly detection, and maritime surveillance.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Dec 14, 2024
- Source ID
- N000142512039
Entities
People
- Katherine Skinner
Organizations
- Board of Regents of the University of Michigan
- Office of Naval Research
- United States Navy