Artificial Intelligence Based Object Detection and Tracking for a Small Underwater Robot

Abstract

Object recognition and tracking is a challenge for underwater vehicles. Traditional algorithm requires a clear feature definition, which suffers from uncertainty as the variation of occlusion, illumination, season and viewpoints. A deep learning approach requires a large amount of training data, which suffers from the computation. The proposed method is to avoid the above drawbacks. The Siamese Region Proposal Network tracking algorithm using two weights sharing is applied to track the target in motion. The key point to overcome is the one-shot detection task when the object is unidentified. Various complex and uncertain environment scenarios are applied to evaluate the proposed system via the deep learning model’s predictions metrics (accuracy, precision, recall, P-R curve, F1 score). The tracking rate based on Siamese Region Proposal Network Algorithm is up to 180 FPS.