Underwater swimming fish mass estimation based on binocular vision

Fish mass is important to optimize daily feeding, control stocking densities, and ultimately determine optimal harvest time. Currently fish mass mainly relies on manual measurements, which is not only time consuming and laborious but also brings stress to fish, and even causes death. However, fish different postures and tail deformation during swimming are challenging for binocular vision to precisely measure mass in unconstrained underwater environment. This study creatively introduces the angle between space vectors to select fish postures, and develop mass prediction model based on length of fish without tail fin. Firstly, three-dimensional coordinates of feature point obtained by convex hull algorithm are calculated by information fusion based on depth images and RGB images from instance segmentation. Secondly, the angle between fish body long axis and camera optical axis is first proposed to select good fish posture for length estimation. Finally, the relationship model between mass and length without considering tail fins is constructed, and subsequently this model has been tested on test datasets. We conducted experiments with Oplegnathus punctatus as an example, and experimental results show that the method can achieve non-contact precise measurement of underwater fish mass, which can be further used to estimate fish mass in densely breeding environment.