An Adaptive Fast-RCNN Method for Fish Monitoring: From an Artificial Environment to the Ocean

Due to the complexity of the marine ecosystem and the limited visibility offered by the underwater medium, the exploration of underwater environments presents numerous challenges. Camera data can provide valuable information about the underwater environment, but it’s often difficult to interpret it accurately. Nowadays, robotics and artificial intelligence advancements are now opening up new opportunities for improving underwater exploration capabilities. The aim of this paper is to develop a system that can identify and follow different elements in the submarine environment with accuracy. The proposed system will establish connections between features that have been extracted from real underwater scenes. In the proposed method, a novel visualization system is designed to enhance the interpretation of submarine environment in order to improve the decision-making capabilities of underwater vessels and autonomous robots. To extract fish characteristics and identify different fish species, an adaptive fast RCNN algorithm will be defined. On the other hand, a Kalman filter will be employed to extract the trajectory of each detected fish. In addition, fish pose in three-dimensional space will also be retrieved. The proposed system was tested using a sophisticated underwater dataset. The experimental outcomes show good progress compared to the most recent state-of-the-art methods.