Underwater Target Recognition and Path Planning Using Otsu-KMeans Segmentation and EM-PF-SLAM with Enhanced A-JPS Algorithm

Abstract

Autonomous Underwater Vehicles (AUVs) face significant challenges in obstacle recognition and path planning due to sonar image noise and uncertain underwater environments. This study proposes a robust target recognition and obstacle avoidance framework integrating forward-looking sonar image processing, probabilistic SLAM, and a fast global path planner. Sonar images are first segmented using Otsu thresholding and refined through K-means clustering to extract obstacle features. These features are then fused with odometry and inertial data using an Expectation-Maximization (EM) based data association module and Particle Filter (PF) for posterior state estimation, enabling accurate SLAM under sonar uncertainty. For navigation, an improved A*-JPS algorithm is applied to achieve time-efficient path planning. Experiments were conducted on the publicly available SeaNet Dataset, which contains diverse acoustic scenes from Northwest Pacific environments. Tests were run on a platform with an Intel i5-10210U CPU and 16 GB RAM, using evaluation metrics including recognition accuraacy, RMSE, absolute trajectory error (ATE), loop closure recall, and path smoothness. Results show that the proposed Otsu-K-means sonar segmentation achieves 99.3% recognition accuracy, outperforming standalone methods by over 25%. The EM-PF-SLAM system achieves an ATE of 0.42 m and a loop closure recall of 95.3%, reducing localization uncertainty by over 40% compared to PF-only baselines. The hybrid A*-JPS planner reduces path planning time to 0.12 s and achieves a smoothness score of 0.85. These findings highlight the method’s suitability for real-time, high-precision AUV operations in complex acoustic environments.