The visual optical guidance is a mainstream solution for the near-end recovery of unmanned underwater vehicles (UUVs), whose accuracy and robustness are susceptible to the limited computational power of onboard devices and interferences of complex underwater environments. In order to address the structural characteristics of rigid optical array targets and the requirements of underwater optical guidance tasks, this study proposes a real-time keypoint detection method for the underwater optical arrays based on the improved RTMPose method. The core contributions are as follows: 1) Establishing a skeleton model for the underwater optical arrays, constructing a dataset for the underwater optical array keypoint detection and adapting the top-down human pose estimation paradigm for the optical guidance tasks; 2) Designing a lightweight backbone network integrating re-parameterization technology and a coordinate attention mechanism. By decoupling the training-inference process and embedding the directional positional information fusion mechanism, the number of parameters and computational complexity are significantly reduced with the compensating ability of performance loss caused by lightweight design, thereby enhancing the model′s robustness against the underwater image degradation, occlusion, and other interferences. Experimental results show that the proposed method achieves an average precision (AP) of 90.8% and a mean pixels error of 0.828 with our dataset. Compared to the baseline model, the proposed model reduces the number of parameters and floating-point operations by approximately 22.8% and 26.6% with the Jetson AGX Orin (64 GB) embedded platform, the corresponding end-to-end inference latency is also reduced by 1.14 ms compared to the baseline. The generalization experiments demonstrate that the method can quickly adapt to different optical array configurations and various water conditions through training, which exhibits the fair versatility and reliability. In conclusion, the proposed underwater optical array keypoint detection method achieves a favorable balance among detection accuracy, speed, and robustness, which meets the precision and real-time requirements for the underwater optical array keypoint detection tasks and provides an effective solution for the high-robustness optical guidance on UUVs and other onboard devices.