<![CDATA[Coal-fired power plants require correct PPE use, but low light in turbine halls, corridors, and nighttime inspection routes limits automated visual monitoring. This study builds and tests a real-time detection-and-tracking pipeline combining a YOLOv9t detector, TensorRT inference, and ByteTrack multi-object tracking for PPE violation monitoring under these conditions. A photometric check of the evaluation footage (mean pixel intensity 117.86/255, 14.68% of pixels below the 64/255 dark-pixel threshold) places it in the dim-illumination band defined by the IEEE 1789 photometric convention, confirming the reported gains were measured under genuinely degraded lighting. Trained on a violation-aware dataset of 11,407 annotated images across eight compliance and violation classes, TensorRT cut mean inference latency from 34.55 ms (PyTorch) to 6.26 ms (FP32) and 5.25 ms (FP16), a paired t-test (t = 13.1223, p 0.001) confirming the engine-level gain, while mAP@50 dropped only marginally (0.9134 to 0.9084). Because TensorRT builds can vary run to run, FP32 and FP16 were each validated across independent builds: FP16 matched FP32 in accuracy and ran 1.32x faster (5.25 vs 6.93 ms, p 0.001), so FP16 was selected for the pipeline below. Extending the comparison beyond one tracker, ByteTrack, DeepSORT, and SORT were benchmarked on identical input from a 128-second, 3,840-frame low-light CCTV clip at the PLTU Paiton gate. ByteTrack reached 82.62 FPS with MOTA 0.7543 (relative to a pseudo ground truth), clearing our 60 FPS real-time threshold (2× the 30 FPS source rate) with the best tracking accuracy of the three; DeepSORT managed only 27.08 FPS despite re-identification embeddings, and SORT (90.29 FPS) produced over 1.3× ByteTrack's identity switches. Across 320 worker identities, mean track duration was 26.74 frames (0.89 s).]]>
