The application of eye movement control of unmanned underwater vehicles is constrained by various factors, one significant reason being that the existing eye movement control methods fail to provide a smooth and natural interactive process for operators. This is due to gaze point jitter and insufficient accuracy and speed in gaze point estimation. Accordingly, this paper proposes an advanced KDE-RF eye movement control method for enhancing the interactive iteration through kernel density and random forest. Specifically, a fusion module is designed to obtain the gaze vector from the operator’s facial image, while a Back-propagation model is then employed to estimate the gaze point coordinates. Furthermore, the kernel density estimation method adopts multiple frames of gaze points and aggregates them into a single frame, which smooths gaze point movement trajectory. The random forest classifier is applied to identifying the UI controls that the operator intends to manipulate, through the pre-determined gaze vectors, thereby guiding the operator to complete interactive tasks more efficiently. Benchmarking experiments demonstrate that the proposed method reduces the average interactive time by 3.09% compared to touch screen interaction and achieved an accuracy of 92.1%. The results validate the proposed eye movement control optimization method simplifies the operational process and has practical application values.