A Lightweight Single-Target Tracking Model for Underwater Sonar Scenarios

Compared with optical images, underwater sonar images suffer from blurry edges, obscure features, and noisy underwater environments, while the small computing power of AUVs further limits the application of large models in underwater environments. To improve the tracking performance of AUV. In this paper, on the basis of siamese network, we designed a lightweight single-object tracking model applied in multibeam forward-looking sonar scenarios, which incorporates the attention mechanism in the feature extraction stage to make the model focus on the target candidate region with high weight, and next the feature fusion module is used to dilute the non-target region features and highlight the features of the region with high similarity, and finally, the output is detected to obtain the target position. The experimental results show that this model can achieve 63% accuracy while keeping the 0.75M parameter count on our sonar dataset.