A Learning-Based Pose Control of Autonomous Underwater Vehicle for Fish Net Inspection in Turbid Water

Abstract

Fisheries are one of the major protein suppliers in the world and more than half of the production is related to aquaculture. In aquaculture, damaged fish net caused by various factor can cause huge losses to fishermen. In order to prevent loss, it is important to detect and deal with the damage quickly. Detecting damaged fish nets using autonomous underwater vehicles (AUVs) may be an efficient and safe solution for avoiding dangers to human divers. Optical camera is needed for more accurate underwater inspection, especially for net inspection, but optical camera suffers visibility degradation with floating particles that cause light attenuation in turbid underwater environments. For getting clear images of the objects to be inspected, therefore, AUVs should come close to the object (e.g. net). In this study, a novel learning-based pose control method of AUV for fish net inspection in turbid water, based on appropriate image features, such as mean gradients over the entire image, is proposed. In the control process, target value should be selected by an operator. To make the system more autonomous, a desired pose (distance) target value can be set by a convolutional neural network (CNN) model that is trained in offline by supervised learning methods. The proposed method can make the AUV maintain a somewhat constant pose with respect to a fish net, which is good enough to acquire clear images of the net to check whether a part of the net is damaged or not under turbid water. Experimental results in three environments, namely, virtual, swimming pool, and real fish farm environments, demonstrated the effectiveness of the proposed methods.