Real-Time Missing Marker Recovery for Optical Surgical Tracking via Dual-Attention Fusion

Optical surgical tool tracking fails when infrared markers are partially occluded and the number of visible markers becomes insufficient for rigid-body pose solving. To mitigate this problem, we study short-horizon marker prediction for binocular optical tracking and formulate it as a unified task covering both pure history-driven prediction and online recovery with partial posterior observations. We propose a dual-attention real-time fusion model that combines a history-driven structural prior with step-wise posterior correction from visible neighbor markers. The architecture consists of a historical encoder, a real-time encoder, a structure-aware module, an interaction-aware module, and a posterior decoder, enabling the model to exploit both temporal dynamics and rigid-body geometry in a causal manner. Experiments on an in-house optical tracking dataset with simulated occlusions show that, under the pure prediction setting without visible future neighbors, the proposed method remains competitive with representative trajectory-forecasting baselines. When one or two future neighbors remain visible, it consistently outperforms representative missing-marker recovery methods across all prediction horizons. Averaged over T1-T5, the proposed method achieves 1.68 mm and 1.38 mm RMSE under N_vis = 1 and N_vis = 2, respectively, corresponding to 14.1% and 18.8% reductions relative to the strongest recovery baseline, HGNN. These results indicate that historical motion modeling alone is insufficient for accurate long-horizon recovery in this task, whereas explicit fusion of real-time structural observations is critical for maintaining marker continuity and supporting robust downstream tool pose estimation.