Dynamic construction behavior recognition through domain-adaptive multimodal evidence fusion

Construction behavior recognition must remain reliable when acquisition conditions or participants differ from the training data. Existing adaptation can admit unsupported target samples, while multimodal fusion can become overconfident when streams are degraded, missing, or temporally misaligned. We propose reliability-coupled domain-adaptive evidence fusion (RDAEF), which derives a shared diagnostic state from signal quality, class-prototype distance, evidential uncertainty, open-set risk and, when multiple streams are available, temporal lag and opinion conflict. The shared state is mapped to a target-admission weight during training and modality-specific evidence discounts during inference. Evaluation uses two complementary configurations: the Construction Meta Action (CMA) benchmark supplies RGB clips for cross-condition adaptation, whereas the Scaffolding Worker IMU Time-series (SWIT) benchmark supplies synchronized pose and inertial sequences for cross-participant adaptation and multimodal fusion. RDAEF achieved target-domain macro-F1 scores of 0.798 on CMA and 0.783 on SWIT, exceeding the strongest unsupervised baselines by 3.4 and 5.9 percentage points, respectively. Removing conflict discounting increased the risk–coverage area (RCA) from 0.118 to 0.198, indicating poorer selective decisions.