Underwater Target Tracking Based on Uncertainty-Inspired Image Enhancement

The task of Underwater Visual Object Tracking (UVOT) not only requires dealing with the common challenges in outdoor tracking but also faces many unique difficulties specific to the underwater environment, including but are not limited to, optical degradation and scattering, uneven illumination, low visibility, and hydrodynamics. In these scenarios, directly applying a large number of traditional outdoor scene object tracking methods directly to underwater scenes inevitably leads to performance degradation. To address the above issues, first, an Underwater Image Enhancement (UIE) module inspired by uncertainty is introduced, aimed at specifically improving the quality of underwater images. This method decomposes UIE into distribution estimation and consensus processes and introduces a new probability network to learn the enhancement distribution of underwater images, thereby addressing the bias problem in reference images. These are subsequently applied to an attention-based feature fusion network to propose a target tracking algorithm, called UTransT. The feature fusion network combines self- and cross-attention mechanisms to effectively fuse template and search region features. The experimental results show that on the UTB180 dataset, the success rate of UTransT is 0.8 percentage points higher than that of MixFormer, with the best performance in the comparison algorithm, and normalization accuracy is nearly 1.9 percentage points higher. On the VMAT dataset, the success rate is 1.2 percentage points higher than that of the best-performing Masked Appearance Transfer (MAT) algorithm, with 1.5 percentage points higher normalization accuracy. Moreover, UTransT facilitates real-time tracking at 65 frames per second. These experimental results validate the effectiveness and feasibility of the proposed algorithm in underwater object tracking tasks.