Real-Time Detection and Counting Method for Distant-Water Tuna Based on Improved YOLOv10n-EMCNet

Area of Science:

• Computer Vision
• Marine Biology
• Robotics and Automation

Background:

• Automated catch monitoring is crucial for distant-water tuna fishing but faces challenges like illumination variation, clutter, and occlusion.
• Existing methods struggle with real-time, accurate detection and counting of tuna on deck operations.

Purpose of the Study:

• To develop an improved lightweight object detector for real-time tuna detection and counting.
• To create an end-to-end pipeline for automated catch monitoring in challenging marine environments.

Main Methods:

• Proposed YOLOv10n-EMCNet detector with backbone, neck, and head enhancements (ESC-based C2f, SMFPN, CAFM).
• Integrated detector with DeepSORT and ROI-based de-duplication for a detection-tracking-counting pipeline.
• Conducted controlled comparisons of tracking algorithms (DeepSORT, ByteTrack, OC-SORT) and shipboard video validation.

Main Results:

• YOLOv10n-EMCNet achieved 94.84% mAP@0.5 and 91.77% recall on the tuna dataset with 6.5 GFLOPs.
• DeepSORT demonstrated the best balance of accuracy, stability, and efficiency for tracking.
• The proposed pipeline achieved 91.4% average counting accuracy with low error rates (8.62% relative, 1.25 fish absolute) at ~30 FPS.

Conclusions:

• The YOLOv10n-EMCNet based pipeline shows significant promise for automated tuna monitoring in real-world shipboard conditions.
• The method effectively addresses challenges of illumination, clutter, and occlusion for reliable catch monitoring.
• This work provides a foundation for advancing automated systems in fisheries management and research.