Underwater target detection using deep learning

Research in marine robotics aims developing Autonomous Underwater Vehicles (AUVs) that are capable of identifying their environment and performing their tasks independently. This requires AUVs to recognize targets useful to their mission. Typically acoustic and optical sensors are used for this purpose. In this study, we employed a camera and deep learning techniques for target detection. We experimented with the latest YOLO architecture YOLOV5 which has impressive performance. The models have been trained with acquired underwater images of the target objects. Training is done using Google-Collab over the cloud to speed up training process. Trained models are then executed on board computer of AUV which is a Raspberry-Pi4 with coral USB accelerator. Target detection task validated with AUV in actual environment. Using rather inexpensive hardware, we are able to achieve a high target detection rate and accuracy.