Multi-Object Tracking of Dense and Uniform Targets for NZ Mussel Farms Using Computer Vision and Genetic Programming Techniques

<p><strong>New Zealand’s aquaculture sector, particularly mussel farming, plays an important role in the country’s economy and export industry. However, its growth and expansion is impeded by operational challenges in the monitoring of mussel farming equipment, particularly the buoyancy devices known as mussel floats/buoys. Their primary function is to provide flotation for the submerged longlines from which mussels are suspended as they grow. Conventional manual inspections are costly and inefficient, making automated tracking solutions increasingly attractive. This thesis addresses the site-specific challenges by developing a new multi-object tracking (MOT) framework tailored for tracking mussel floats in mussel farm images. The research explores the integration of traditional computer vision and Genetic Programming (GP) techniques to facilitate robust and reliable tracking in dynamic and unpredictable aquafarm environments where deep learning methods often underperform.</strong></p><p>This research delivers three major contributions. Firstly, a new training-free MOT pipeline based exclusively on traditional computer vision methods was established to provide a robust performance benchmark for mussel float tracking. Secondly, this thesis introduces a GP-based multi-object detection approach to identifying mussel floats in images, featuring a new hierarchical 3-Tree GP program representation that significantly improves detection accuracy and reduces false positives. Lastly, this thesis introduces a GP-based multi-object matching approach to reliably matching large numbers of visually similar mussel floats across frames.</p><p>The findings demonstrate that GP can effectively enhance both multi-object detection and matching performance in MOT, outperforming conventional and deep learning-based methods. The proposed automated MOT approaches reduce the need for manual inspections, enhance operational efficiency, and lay the groundwork for future research into scalable aquaculture monitoring systems. Through the application of artificial intelligence, this thesis provides a practical solution that supports the advancement of aquaculture in New Zealand, with positive implications for both the economy and the environment.</p>