Research on Image Localization Algorithm for Underwater Dynamic Target

In the underwater robot operation task, the position and form of the dynamic operation target may change on any occasion. To meet the requirements of high real-time and high precision in underwater living targets and other dynamic target grasping operations, this study proposed an image localization algorithm for underwater dynamic targets. The algorithm takes the target detection window generated by the deep learning model as the initial search window for the correlation filter, effectively narrowing the tracking range. Then, it adopts the correlation filter techniques to calculate the target patches based on the target search window to meet the real-time demand. Through the introduction of the accuracy judgment mechanism, it iteratively updates the two-dimensional form-centered coordinates of the acquired target to improve the positioning accuracy. Ultimately, this approach combines binocular stereo vision techniques to achieve real-time and precise three-dimensional localization. The underwater experimental results indicate that, compared to similar algorithms with equivalent positioning accuracy, the time required for three-dimensional localization of dynamic targets by the same algorithm is 45 ms. In comparison, the time consumed by the algorithm of this study for three-dimensional localization of dynamic targets is 24 ms, which effectively improves the real-time and accuracy of underwater dynamic target localization.