STTA-SlowFast: Spatial temporal triplet attention SlowFast for pilot action recognition

For competency-based flight training assessment, this paper introduces a Spatial-Temporal Triplet Attention (STTA) module into the SlowFast framework to explicitly model cross-dimensional interactions between the temporal dimension [[EQUATION]] and channel/height/width [[EQUATION]], enhancing fine-grained action representation with minimal overhead. A cockpit pilot operation dataset is constructed with 614 keyframes across 8 classes from 45 pilots, and a keyframe-centered symmetric sampling strategy with temporal jitter is adopted. On keyframe-based action recognition, STTA-SlowFast achieves 94.84% Top-1 accuracy on our dataset, representing a 5.81 percentage point improvement over the baseline. This gain is achieved with only 0.008 M additional parameters and 0.49 G additional FLOPs. Ablation and Grad-CAM (Gradient-weighted Class Activation Mapping) results demonstrate complementary effects across the three branches. The analysis shows that continuous spatio-temporal attention bands effectively mitigate symmetric confusions (e.g., thrust-lever forward vs. backward) and improve the stability of recognizing multi-stage, small-amplitude operations (e.g., flap lever, instrument knobs), providing a feasible path for automated logging of pilot operational behavior.