AUV pose correction via underwater object recognition using synthetic data

Accurate localization of autonomous underwater vehicles (AUVs) is challenging because inertial measurement units (IMUs) and Doppler velocity logs (DVLs) accumulate drift during long-duration missions. To address this challenge, this study explores whether synthetic sonar imagery can provide transferable landmark cues for real-world mapping, thereby reducing dependence on costly and limited field datasets. Using an Unreal Engine 5-based FURo-sim framework, we generated forward-looking sonar (FLS) data and automatically annotated high-intensity returns and their acoustic shadows. These annotations are used to train a YOLO11 detector exclusively on synthetic data, and the centroids of the resulting bounding boxes are interpreted as landmark observations. The landmark detections are then fused with DVL/IMU dead reckoning within a two-dimensional pose-graph optimization framework to refine both the vehicle trajectory and the resulting sonar mosaics. The proposed approach is validated using real coastal-sea data, with ultra-short baseline (USBL) acoustic positioning serving as a proxy for ground truth. Experimental results show that the method reduces dead-reckoning localization error by $$15.4\%$$ , demonstrating the effectiveness of the synthetic data-driven detector-based pose-graph formulation for real-world AUV navigation.