Inspection Operations in Fish Net Cages through a Hybrid Underwater Intervention System using Deep Learning

Abstract

Net inspection in net cages is a daily task for divers at the fish farms. This task represents a high cost for fish farms and is a high risk activity for the divers. Net cages are basically big structures with a depth of more than 20m and around 25m diameter. The total inspection surface can be more than 1500 $m^{2}$, which means that this activity is time-consuming. Considering that divers have limited time underwater, this activity represents a significant area for improvement. Additionally, a net pen is a harsh environment with hundreds of fish swimming, fish morts and ocean currents as some of the phenomena to consider. Some works have addressed this problem using underwater robots equipped with sensors such as USBL or DVL, and applying different control theories to offer a solution to this problem.

A platform for net inspection is proposed in this Thesis. This platform includes a surface vehicle, ground station and an underwater vehicle embedded with artificial intelligence and control trajectories. The underwater robot used is the BlueROV2 on its heavy configuration, some localization techniques are used to control the position of the robot such as a monocular camera at the surface vehicle using an ArUco code and object detection. Computer vision is also implemented in this work, a Convolutional Neural Network was trained in order to predict the distance between the net and the robot. Finally, some experimental results about the hole detection and position algorithm, the net distance estimation and the inspection trajectories are also presented that demonstrate the robustness, usability, and viability. The experimental validation took place in the CIRTESU tank, which has dimensions of 12x8x5 meters, at Universitat Jaume I.