Underwater obstacle avoidance for AUV using forward-looking sonar based on deep learning: Method and experiment

An integrated solution for autonomous underwater vehicle (AUV) obstacle avoidance using forward-looking sonar (FLS) is presented. This approach addresses critical underwater navigation challenges, particularly high false-alarm rates in complex acoustic environments and reactive avoidance limitations under maneuverability constraints. A detection framework is developed using native rectangular grayscale sonar data reconstructed directly from the acoustic stream to preserve acoustic signatures and avoid distortions inherent in conventional fan-shaped transformations or pseudo-color processing. Furthermore, a geometry-based annotation strategy is introduced to enhance model generalization across various underwater targets. To mitigate unnecessary maneuvers triggered by acoustic artifacts that cause significant trajectory deviations for underactuated AUVs, a confidence-based dynamic parameter adaptive tracking algorithm is proposed. This algorithm adaptively adjusts matching criteria to effectively suppress transient false positives. Additionally, an improved real-time avoidance line-of-sight (RA-LOS) guidance law is developed, incorporating adaptive obstacle buffer sizing to ensure feasible maneuvers within the vehicle’s motion constraints. The integrated system was validated through comprehensive lake trials. Experimental results demonstrate that the proposed tracker achieves a 49.5% suppression rate for low-confidence false positives, while the guidance law ensures a 100% success rate in avoiding both artificial obstacles and natural boundaries. This study provides robust experimental validation of the practical viability of deep learning-based perception and adaptive guidance for real-time AUV obstacle avoidance.