The SeaClear2.0 robotic system’s underwater litter detection component uses neural networks trained on image and video data.

In this video, we explore the network’s predictions using video data it has never seen before. The network can successfully detect various litter items and marine life species, but there’s still room for improvement. Our colleagues at TU Delft are currently expanding the dataset used for training the network with new images and videos for more reliable object detection.

Yet, underwater environments can be rather challenging when it comes to object detection – images are often blurry due to turbid water or suboptimal lighting conditions.

As such, TUDelft is currently training the neural network using a combination of camera and sonar images, which could further improve the system’s predictions.