Optical flow prompts distractor-aware siamese network for tracking autonomous underwater vehicle with sonar and camera videos.

Underwater moving target tracking with sonar device is very challenging. Due to complex underwater environment, sonar acquired videos have domain-specific challenges like texture-similar backgrounds and variable visual attributes, impairing the performance of existing tracking models. To address these issues, we propose a visual tracking model named OFDTrack to overcome the limitations of existing methods. First, for moving target in sonar videos, a re-capture paradigm is proposed to search potential motion areas from rough prompts of the optical flow field, thus maintaining tracking in the event of target loss. Furthermore, for the variations in visual features, a dynamic update template with constraints is designed, enabling the model to maintain target tracking when significant deformations occur. In order to eliminate the disturbance caused by the wake generated by propellers of Autonomous Underwater Vehicle, an auxiliary tracker is embedded to formulate the motion pattern of moving target and its prediction corrects the deviation caused by abnormal displacement. Finally, in real-world tests, we collect video of the same moving Autonomous Underwater Vehicle in three test scenarios using sonar and camera mounted on Unmanned Aerial Vehicle respectively, and verify the effectiveness and robustness of our method through comparisons with other tracking models.