Simulating Sonar Data for Cost-Effective Underwater Place Recognition

The increasing need to explore and monitor underwater environments has led underwater robotics to an impressive growth in recent decades; object detection, environmental mapping, underwater exploration, and inspection missions are tasks routinely performed by autonomous underwater vehicles (AUVs) in modern underwater operations. However, one of the biggest challenges in autonomous underwater navigation is the capability of AUVs to accurately localize themselves since common positioning systems are not available and optical sensors are strongly limited by environmental characteristics such as light absorption and water turbidity. A solution to this problem is represented by sonar systems, which are widely used for underwater perception as they are not affected by these limitations. This thesis was developed as a part of a project collaboration between University of Padua and SAIPEM S.p.A., and addresses the Underwater Place Recognition (UPR) task by generating a large-scale synthetic dataset composed of labeled multi-beam sonar images, acquired through DAVE simulator emulating the behavior of the Tritech Gemini 720i multibeam sonar mounted on SAIPEM's FlatFish AUV. Generated dataset is then used to train and evaluate a deep learning method for UPR. It extracts global descriptors from sonar images using a neural architecture based on the ConvNeXt, paired with a customized decoder. The model is trained using a novel Batch Contrastive Learning strategy that exploits the oriented Intersection over Union (o-IoU) as a metric to guide the learning process, ensuring that descriptor distances reflect the spatial and angular similarity of sonar viewpoints. Performances are evaluated on both synthetic and real sonar data to assess the generalization capabilities of the system. Furthermore, alternative architectures are tested to investigate the influence of different design choices on UPR performances. Preliminary results demonstrate the effectiveness of synthetic data for training deep learning-based place recognition systems, offering a promising direction for improving underwater localization in future AUVs deployments.