Fish length is a key parameter to assess fish stocks, it helps management and brings economic benefits in the context of aquaculture. The collection of information on the size of fish is essential in many stages of production, such as (a) determining the ideal moment of capture, (b) estimating a series of parameters at the population level and studying its evolution (volume, weight, sex or fat content); (c) stock assessment during creation; (d) classification and separation of individuals by size; (e) optimization in food distribution; (f) controlling individuals’ growth; (g) forecasting market value.
To obtain length measurements, a fish ruler is usually used. A cumbersome task that, when applied to living individuals, can induce considerable stress, increasing the risk of damage or hindering their growth. Computer vision is one of the most used non-contact tools for measurement, being fast, consistent and repeatable. However, its use in aquatic environments is limited by the high cost, the difficulty of calibrating the system underwater and the complexity of implementation.
This presentation proposes a low-cost easy-to-use vision system that can obtain measurements on live fish within the aquatic environment, without the need for calibration and a demanding image analysis service. In underwater conditions, the captured images of a vision system undergo a refraction effect, as light rays cross a multiplicity of media (e.g. water, air, lens, glass). This effect increases with the angle of incidence of the light rays on the contact surface between two media. Consequently, the usual three-dimensional vision tools for a stereo camera system (in mid-air) are no longer applicable, invalidating the epipolar geometry associated with the stereo system. It becomes necessary to calibrate the system exhaustively, as the matching map is heavily distorted, a task difficulted by the need to be performed underwater.
The present work implemented a compact stereo vision system and developed an algorithm that estimates the correct length of fish, based on the variation of the angle of incidence of the light rays in the water. Given some structural conditions such as a short baseline (on the order of 100mm) and an ROI (Region of Interest of the FOV, measured in degrees) of less than 60º, the system measures fish with an error of less than 1%, using a default calibration. The short baseline allows to have a compact system and benefits the length estimation algorithm. A set of experiments were performed with real fish, working robustly for a set of orientations of the fish (even when the tail and nose are on different distances to the cameras). The photos on the left show two views of the system that integrates a computer and a stereo pair; on the right, two images captured, and a measurement of an individual are presented