Accurate perception of a diver’s position and orientation by Autonomous Underwater Vehicles (AUVs) is essential for effective human–robot collaboration in underwater environments. However, conventional position and orientation estimation methods that combine deep learning with Perspective-n-Point (PnP) algorithms are primarily designed for rigid objects. In contrast, divers exhibit highly variable postures underwater, with no fixed configuration. To address this limitation, this paper proposes a framework for estimating the six-degree-of-freedom (6-DoF) position and the orientation of a diver. In addition, a novel network architecture, termed “VideoPose5CH,” is proposed. In the proposed framework, temporal sequences of 2D joint coordinates are provided to VideoPose5CH, which then outputs the 3D joint coordinates of the current frame as well as the corresponding refined 2D joint locations. Subsequently, the diver’s 6-DoF position and orientation relative to the camera are further recovered via a PnP algorithm. To mitigate the scarcity of underwater 3D human pose datasets, a land-based 3D human pose dataset augmentation strategy tailored to underwater conditions is further proposed. With this strategy, diver pose estimation accuracy is improved and the robustness of the proposed method across diverse scenarios is enhanced. Experimental results demonstrate that the proposed method can stably estimate the 6-DoF position and orientation of the diver within a distance range of 2.643 m to 11.477 m. The average position errors along the three axes are 7.33 cm, 4.04 cm, and 27.15 cm, respectively, while the average orientation errors are 6.96°, 8.47°, and 2.62°.

Underwater image degradation and occlusion make diver keypoints less visually discernible, leading to substantial localization errors in the keypoint pseudo-labels generated by existing semi-supervised methods. These errors can then accumulate and propagate during consistency learning. Moreover, these methods do not fully exploit the human structural information available in unlabeled images. To address these limitations, a limb-heatmap-guided bidirectional cross-representation consistency method is proposed for semi-supervised learning. A keypoint network and a limb network are employed to learn point-level keypoint responses and region-level limb-structure responses, respectively. For labeled data, limb labels are generated automatically from keypoint annotations and a predefined skeletal topology, without additional annotations. For unlabeled data, intra-representation consistency learning is performed for both networks between the easy-augmentation and hard-augmentation views, and bidirectional cross-representation supervision is applied between the two networks. In the keypoint-to-limb path, keypoint predictions are converted into limb pseudo-labels to supervise the limb network; in the limb-to-keypoint path, limb heatmaps reconstructed from keypoint predictions are compared with the predictions of the limb network to constrain the keypoint configuration. During model inference, only the keypoint network is retained, with no additional inference cost. On the independent test set of a single underwater diver dataset constructed for this study, the proposed method achieves a mean PCKh@0.5 of 80.07% using all labeled training data and outperforms the comparison methods at labeled-data proportions of 25%, 50%, 75%, and 100%. These results demonstrate that region-level limb structure provides effective complementary supervision for point-level keypoint localization within the evaluated setting. Broader applicability requires further validation.