Underwater Sonar Visual Tracking with Obstacle Solving Method

In the field of visual target tracking, the efficient convolution operators (ECO) algorithm has performed very well in accuracy and efficiency. Unlike optical images, underwater sonar images exist the problems of low resolution, high similarity between foreground and background, no obvious texture features. And the underwater moving organisms will deform frequently in a short time. We refer to these problems collectively as obstacles. Because of these obstacles, if an excellent target tracker is directly applied to underwater sonar image sequences, it will not achieve ideal results. Therefore, this paper proposes an underwater sonar visual tracker (USVT) capable of stable tracking. Based on the ECO tracker, first of all, for the problem of high similarity between the foreground and background, we change the original ECO model update method, and make it update under the condition of high-confidence tracking results to repress model drift. Second, we add a deformation suppression model that can repress the change of the response map, which lead to a better effect on targets that are frequently deformed at low resolution. Experiments were performed on the underwater sonar image datasets (USDT). Compared with the algorithm before, it proved that the new method proposed in this paper has better tracking performance for underwater sonar image sequences.