Automatic fish measurement using a camera and a 3D sensor applied to a long-term experiment

Abstract

The fish monitoring effort has been increasing over the past years, due to conservation and management requests demanding more accurate data and consequently raising costs. This is an important challenge especially for remote and disperse locations where fish sampling poses unbearable costs, leading to limited spatial sampling schemes, limited data on rare and occasional landed species, and erroneous and biased observations. In this article, we propose a new autonomous system that can be installed on monitoring spots or on board fishing vessels, which is able to remotely acquire all the landed or captured fish and measure it automatically without any physical interaction. The system uses (i) a camera and a 3D sensor to obtain a complete XYZ reading of the fish and (ii) a convolutional neural network, trained for a representative set of species to detect and measure the individuals visible in a box. The system was validated in real conditions, using continuous observations of the landed fish in three islands of the Azores Archipelago, for 2 years. The aim of this article is to provide an overview of the measuring system and an analysis of the sampled data, by comparing the results of the proposed method with the traditional sampling methodology for a given period.