03/19/2026
By Danielle Fretwell

The Francis College of Engineering, Department of Electrical and Computer Engineering, invites you to attend a Master's Thesis defense by Quoc Huy Huynh on: "Sonar-Centric Perception for Underwater Robots: Sim-to-Real Object Detection and Fourier-Based Scan Registration."

Candidate Name: Quoc Huy Huynh
Degree: Master’s
Defense Date: Wednesday, April 1, 2026
Time: 3:15 – 5:15 p.m.
Location: Southwick 240

Committee:

  • Advisor: Paul Robinette, Ph.D., Department of Electrical and Computer Engineering, University of Massachusetts Lowell
  • Maru Cabrera, Ph.D., Miner School of Computer and Information Sciences, University of Massachusetts Lowell
  • Jean-Francois Millithaler, Ph.D., Department of Electrical and Computer Engineering, University of Massachusetts Lowell

Abstract:
Autonomous underwater robots rely heavily on sonar sensing for perception and navigation in environments where optical visibility is limited. However, sonar imagery presents significant challenges for robotic perception, including low resolution, speckle noise, and large variations across sensor modalities. These issues complicate the application of modern machine learning and registration techniques that have been successful in optical domains. This thesis investigates methods for improving sonar-based perception for underwater robots through both learning-based approaches and frequency-domain registration techniques.

First, we study the problem of object detection and tracking in sonar imagery under sim-to-real transfer. Deep learning models are trained on synthetic sonar data generated in the HoloOcean simulator and fine-tuned using real-world datasets collected with two sonar modalities: the Ping360 scanning imaging sonar and the Oculus M750d multibeam sonar. Using YOLOv8 for detection and SiamRPN++ for tracking, we evaluate how model performance varies across sensor modalities. Experimental results show that models pretrained on synthetic multibeam sonar generalize well to real multibeam sensors but struggle with scanning imaging sonar due to differences in signal structure and noise characteristics. These findings highlight the importance of modality-aware training strategies for sonar perception systems.

Second, we introduce One-Ring Correlation Alignment (ORCA), a fast spectral method for estimating the relative rotation between two sonar scans. ORCA operates in the Fourier domain by sampling angular signals along radial frequency rings of the magnitude spectrum. Since not all frequency rings provide equally informative directional structure, the method estimates ring reliability using variance-based statistics and aggregates them into a weighted rotational descriptor. The relative rotation between scans is then recovered through FFT-based circular correlation. Experiments on both simulated and real sonar datasets demonstrate that ORCA improves rotational alignment accuracy while significantly reducing computation time compared to existing spectral registration approaches.

Together, these contributions address key challenges in underwater perception by combining learning-based sonar understanding with efficient spectral registration techniques. The results provide insights into the role of sonar modality in sim-to-real learning-based object detection and tracking and demonstrate how frequency-domain analysis can improve scan alignment for underwater robotic mapping and navigation.