Maritime man-overboard search based on MOB-Detector with modulated deformable convolution and bi-directional feature fusion network
IntroductionMaritime transport is vital for global trade and cultural exchange, yet it carries inherent risks, particularly accidents at sea. Drones are increasingly valuable in marine search missions. However, Unmanned Aerial Vehicles (UAV) operating at high altitudes often leave only a small porti...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Frontiers Media S.A.
2025-06-01
|
| Series: | Frontiers in Marine Science |
| Subjects: | |
| Online Access: | https://www.frontiersin.org/articles/10.3389/fmars.2025.1547747/full |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | IntroductionMaritime transport is vital for global trade and cultural exchange, yet it carries inherent risks, particularly accidents at sea. Drones are increasingly valuable in marine search missions. However, Unmanned Aerial Vehicles (UAV) operating at high altitudes often leave only a small portion of a person overboard visible above the water, posing challenges for traditional detection algorithms.MethodsTo tackle this issue, we present the Man-Overboard Detector (MOB-Detector), an anchor-free detector that enhances the accuracy of man-overboard detection. MOB-Detector utilizes the bi-directional feature fusion network to integrate location and semantic features effectively. Additionally, it employs modulated deformable convolution (MDConv), allowing the model to adapt to various geometric variations of individuals in distress.ResultsExperimental validation shows that the MOB-Detector outperformed its nearest competitor by 8.6% in [Metric 1 AP50] and 5.2% in [Metric 2 APsmall], demonstrating its effectiveness for maritime search tasks. Furthermore, we introduce the ManOverboard Benchmark to evaluate algorithms for detecting small objects in maritime environments.DiscussionIn the discussion, the challenge faced by the MOB-Detector in low-visibility environments is discussed, and two future research directions are proposed: optimizing the detector based on the Transformer architecture and developing targeted data augmentation strategies. |
|---|---|
| ISSN: | 2296-7745 |