Ritzau, LindaLindaRitzauBaatar, GanzorigGanzorigBaatarKarimanzira, DivasDivasKarimanziraRauschenbach, ThomasThomasRauschenbach2025-08-182025-08-182025https://publica.fraunhofer.de/handle/publica/49065910.1109/OCEANS58557.2025.11104292This paper presents a real-time object detection and tracking algorithm for forward-looking sonar image sequences to enhance navigation and obstacle avoidance algorithms for autonomous underwater vehicles. Reliable underwater navigation, which is essential for underwater operations, relies on accurate knowledge of the surrounding environment. Due to strong absorption of signals, underwater vehicles are widely equipped with forward-looking sonar sensors, which provide detailed, wide-ranging environmental information. The sonar data serve as a stable foundation for object detection and tracking tasks in autonomous underwater vehicle navigation. However, images from these acoustic-based sensors are often compromised by interference from feature distortion, reverberation, and en-vironmental noise. These distorted features and acoustic noise present significant challenges for detecting and tracking objects in acoustic images. To address these issues, the developed method combines object detection based on visual features in sonar images with vehicle motion information obtained from navigation devices and sensors. The method is based on a deep learning enhanced underwater object detection algorithm involving range-setting and sliding window preprocessing. The additional vehicle motion data significantly enhances precision, recall and processing time of this object detection algorithm. Building on this specific underwater object detection, our tracking method proposes a combination of the detection in sonar images and vehicle movement information. The proposed object tracking algorithm improves object detection by calculating the searched object's localization using navigation and vehicle movement information if the object detection model does not successfully localize the searched object. This allows the algorithm to maintain continuous track of objects even when they are temporarily not visible on sonar images. The combinatory procedure makes the algorithm robust against common sonar image issues, such as noise or low resolution, and enables reliable tracking of static objects in sonar images for subsequent navigational tasks.enDeep learningAutonomous underwater vehiclesSonarObject detectionSonar navigationSearch problemsReal-time systemsObject trackingReliabilitySonar detectionUnderwater Object DetectionUnderwater Object TrackingSonar ImageForward Looking SonarAutonomous Underwater VehicleRemotely Operated VehicleDeep LearningReal Time Object Detectionconference paper