DRAF-Net: Quality-Aware RGB–Skeleton Temporal Fusion for Driver Risk-Behavior Recognition

Occlusion in vehicle cabins and pose-estimation errors can reduce the reliability of skeleton information. Conventional fusion methods may propagate unreliable evidence when the RGB, joint, and bone streams produce conflicting predictions. We therefore propose DRAF-Net, a Dual-stage Risk-Aware Fusion Network. The Pose-Quality-guided Risk Bottleneck Temporal Adapter (PQR-BTA) regulates information exchange among the three streams using 12 skeleton-quality measures, prediction confidence, and cross-stream disagreement. Risk Correction Protection (RCP) then uses the baseline prediction as a reference and bounds changes to the high-risk probability and four-level classification scores. Experiments used three mutually exclusive subject splits from Drive&Act, with three random seeds per split and a shared protocol for input features, training, and validation-based model selection. Across the three subject splits, DRAF-Net achieved a high-risk AUPRC of 96.16±2.44%, exceeding the strongest comparison method by 3.13 percentage points. At the validation threshold targeting a FAR of approximately 5%, Recall reached 89.18±5.48%, an improvement of 5.82 points over the strongest comparison method. The deviations are sample standard deviations across the three split-level seed means. These results show that DRAF-Net organizes complementary RGB and skeleton evidence effectively while supporting both the ranking and detection of high-risk samples.