Parallel Attention Mechanism Based Multi-feature Fusion for Underwater Object Tracking

Recently, siamese network-based trackers have achieved great success, however underwater object tracking has been rarely studied. In underwater environments, the severe deformation, rapid movement, and complex background interference of objects often lead to low accuracy in underwater object tracking. To address the above challenges, we propose an underwater object tracking method based on siamese networks. The proposed parallel attention module facilitates the aggregation of similar semantic features from different positions and promotes information exchange between the two branches, enhancing the feature expression capability between channels in each branch. Moreover, the multi-scale feature fusion module effectively integrates features from various levels to adapt to changes in the target’s appearance. Finally, comprehensive experiments were conducted on the OTB100, VOT2018, and underwater dataset UT40, demonstrating the method has good performance in underwater object tracking.