A new multi-object tracking pipeline based on computer vision techniques for mussel farms

Area of Science:

• Aquaculture technology
• Computer vision applications
• Image processing

Background:

• Mussel farming is vital to New Zealand's economy and communities.
• Current mussel float tracking is labor-intensive, relying on manual boat trips.
• Automating float tracking presents challenges due to identical targets and varied image conditions.

Purpose of the Study:

• To develop an automated computer vision pipeline for detecting and tracking mussel floats.
• To alleviate the labor burden on mussel farmers through technological integration.
• To address the challenge of tracking numerous identical objects in diverse aquatic environments.

Main Methods:

• A novel computer vision pipeline comprising float detection, description, and matching.
• Utilizing image processing operators for robust float detection across various sizes.
• Employing a unique descriptor based on neighbor positions for float identification.
• Implementing image registration for accurate float matching between adjacent frames.

Main Results:

• The proposed pipeline achieved 82.9% MOTA (Multiple Object Tracking Accuracy).
• This performance is 18% higher than existing deep learning-based methods.
• The system operates effectively without requiring model training.

Conclusions:

• The developed computer vision pipeline offers an efficient and automated solution for mussel float tracking.
• This technology significantly improves upon current methods, offering higher accuracy and reduced labor.
• The approach is suitable for real-world application in mussel farming, particularly in environments like New Zealand's Marlborough Sounds.