Underwater Target Tracking Based on Siamese Fully Convolutional Network

The underwater acoustic detection environment is complex and subject to significant noise interference, leading to high similarity between underwater acoustic image speckles and background images. This similarity makes it difficult for target tracking algorithms to accurately obtain histogram confidence estimates, resulting in rough and discontinuous tracking trajectories. To address this issue, a method for tracking underwater acoustic image speckle targets based on Siamese fully convolutional network is proposed. The histogram confidence estimates for the target template and search area in underwater acoustic images are determined, and a self-restarting mechanism for the learning rate is simulated, with the learning module initialized by given correct annotations. The predictive learning module is updated based on the confidence estimates, effectively distinguishing the speckle target to be tracked from similar targets and achieving precise tracking of the speckle target. Experimental results show that our method produces smooth and continuous tracking trajectories for underwater acoustic image speckle targets, exhibits good resistance to noise interference, and has high accuracy and robustness.