Domain-Robust Temporal Action Detection for Power Grid Operation Videos via Consistency-Guided Domain Modeling

Temporal action detection (TAD) should remain reliable when videos are collected by different sources, even if the action vocabulary is unchanged. We study this domain-generalization setting with pre-extracted temporal features and no target-domain videos during training. DRTAD models source-related feature factors, mixes them during training, and applies domain-aware feature decoupling as a training-time transformation. At inference, the learned detector uses the standard detector path without the DRTAD adapter. The main evaluation uses MD-TAD, a cross-dataset benchmark assembled from ActivityNet v1.3, FineAction, HACS, and THUMOS14. DRTAD improves average mAP from 32.7 to 34.0 with BMN and from 42.6 to 44.1 with ActionFormer. An additional industrial evaluation uses a 30% subset of power grid operation videos, with domains defined by acquisition source and temporal difficulty and annotations represented on the processed-clip timeline. These experiments examine transfer beyond public web-video datasets and the contributions of domain modeling, factor mixing, and training-time feature decoupling.