Object Tracking Approach for Catch Estimation on Trawl Surveys

Abstract

In the Norwegian Sea, coordinated multinational surveys are regularly undertaken with the aim of assessing the size and composition of marine life populations - a fundamental practice for ensuring long-term ecological sustainability. The role of trawling in these surveys is pivotal, as it offers a direct, fisheries-independent, sampling method. This direct approach enables an accurate assessment of the abundance and diversity of fish populations, providing a clearer picture of the marine ecosystem's health. However, traditional trawling leads to increased bycatch mortality rate and can hurt biodiversity. Scantrol Deep Vision is a company that focuses on the development of advanced underwater vision technology. They have launched a product known as "Deep Vision" which aims to revolutionize marine research by providing an eco-friendly method for fish sampling and stock analysis without the need to bring the catch onboard. The technology takes pictures of marine life during trawling. Our project attempts to estimate the marine life count and distribution based on these images. Previous work, by Allken et al., on this problem involved fine tuning a RetinaNet model to detect and classify four categories of fish: blue whiting, herring, mackerel, and mesopelagic fishes. They ran the model on images from 20 trawl stations and trained a linear regression model for each species, except mesopelagic fishes, on the resulting object detection count generated form each station and their respective catch count. They used the R-squared metric to quantify how well the regression models fit the data and got the scores 0.74, 0.62, and 0.84 for blue whiting, herring, and mackerel, respectively. Mesopelagic fishes are generally too small to be caught by the trawls and were not part of any regression. In our project, we aim to enhance the precision of estimation on marine life count and species distribution. We employ object tracking to a dataset generated by the same RetinaNet model used in previous studies for object detection. Subsequently, we apply linear regression to the count derived from these tracks. Our most effective model demonstrates promising results, on the 20-station dataset used in previous work, with R-squared scores of 0.84, 0.96, and 0.87 for blue whiting, herring, and mackerel, respectively. These results underscore the potential efficacy of object tracking in addressing this problem.