Vision-based Deep Learning algorithm for Underwater Object Detection and Tracking

Remotely operated (ROV) and autonomous underwater (AUV) vehicles are the efficient tools that are used effectively for different ocean expedition and exploration-based missions. Underwater image processing is one of the challenging tasks in real time and post processing of the recorded underwater visuals. The challenges are due to the water turbidity, scattering, and poor visibility of the light. The acquired underwater visuals should be post-processed using image enhancement techniques for object detection. In this paper, terrestrial based deep learning-based object detection techniques and image enhancement techniques are adapted for underwater object detection in shallow waters. Simulations are carried out for known and pre-trained submerged docking station images for detection and tracking in different water quality scenarios. The proposed technique uses a transfer learning-based object detection algorithm using YOLOV4 (You Only Look Once) model with existing image enhancement techniques and the simulation results show that there is a significant increase in the accuracy of underwater docking station detection using the proposed model compared to the traditional models. The kinematics are computed for the detected object from the dynamic underwater vehicle in different scenarios.