Enhanced Underwater Object Detection in Sonar Images with Range-Setting Embedding and Sliding Window Preprocessing

This paper presents a robust real-time object detection approach utilizing sonar imagery from forward looking sonar systems, aimed at enhancing navigation and obstacle avoidance algorithms for autonomous underwater vehicles (AUVs). Effective autonomous navigation of an underwater vehicle ne-cessitates precise knowledge of the surrounding environment. While cameras are often employed for near field environmental scanning, they encounter significant challenges such as turbidity and light absorption, which can compromise the reliability of visual data. In contrast forward looking sonar provide reliable measurements and detailed environmental information. However, the sonar images pose their own challenges, including low resolution and substantial noise, complicating the accurate identification of visual features. To address these issues, we propose a novel sonar image preprocessing method designed to improve object detection. This method involves systematically moving a predetermined-size window (in meters) across the sonar image, extracting data from various positions to create smaller, fixed-size images. The window size is determined by the sonar device's range setting ensuring that the extracted images will represent a consistent area of the underwater environment. Hence, extracted images consistently represent specific areas of the underwater environment. These preprocessed images are then utilized for training object detection models and as input for detection tasks. This preprocessing step significantly enhances the stability and reliability of target object detection making it compatible with a wide range of existing object detection models.