Dynamic Positioning of an Underwater Vehicle using Monocular Vision-Based Object Detection with Machine Learning

Towards Reliable Localization for Autonomous Unmanned Underwater Vehicle Operations

“…It is widely used for unmanned aerial drone applications [145]- [147], but due to the underwater effects on the image quality, the feature extraction is more complicated underwater [148]. Several methods using neural networks have been proposed for underwater image enhancement to overcome the challenges from underwater effects, including color correction [149].…”

“…Several methods using neural networks have been proposed for underwater image enhancement to overcome the challenges from underwater effects, including color correction [149]. Other methods focus on object detection; some examples of this approach can be seen in [150], where Faster R-CNN is proposed, which differs from its predecessor R-CNN [151] and Fast R-CNN [152], by circumventing the selective search algorithm, which is time-consuming, and makes real-time object detection irrelevant [148]. Also, YOLOv3 has proven useful for object detection; it is the third version of YOLO (You Only Look Once), which is a machine-learning algorithm trained on a large database of labeled images of various objects [153].…”

“…Common for all images used for YOLO is that they have been taken above the surface; therefore, this algorithm needs additional training (transfer learning) with underwater images. In [148], YOLOv3 was used with transfer learning on an additional 5000 underwater images. The algorithm was used for object detection, which was then used for motion control of a BlueROV2.…”