SPCM-Det with Local State Updates for Video Detection in Urban-Rail Protection Zones

Abstract Urban-rail protection zones restrict external activities near rail infrastructure. Visual monitoring can warn operators when hazardous machinery approaches the protected boundary. Existing methods generally strengthen single-frame features, expand temporal context, or transfer representations across acquisition conditions. Their broader spatial and temporal modelling can mix weak target cues with unrelated responses, while global adaptation does not address classes that are sparsely represented in particular scenes. We address these problems with the Scene-Prior Conditioned Mamba Detector (SPCM-Det). It treats short video detection as local temporal state updating at corresponding positions of a multi-scale feature pyramid. The Temporal Mamba Neck (TMN) uses input-conditioned state updates and a gated residual path to control how each observation modifies the \roundrev{local temporal state}. Illumination-Aware Adaptive Normalization (IAN) adapts running daytime and night-time activation statistics to \roundrev{frame brightness}. Scene-Aware Sample Reweighting (SASR) increases the loss contribution of target classes that are underrepresented in a scene. We collected RailRisk-144K, a real-world dataset containing 144,544 video images from protection zones along Guangzhou Metro Line 18. Detection accuracy is measured by average precision (AP) and mean average precision (mAP) at intersection-over-union (IoU) thresholds. Compared with ten baselines on this dataset, SPCM-Det provides average absolute gains of 11.3\%, 12.9\%, and 21.5\% in [email protected], [email protected]:0.95, and \roundrev{night-time [email protected]}. The corresponding maximum gains are 17.4\%, 18.5\%, and 28.4\%.