Abstract
Reliable railway signal light detection remains challenging because distant signals occupy only a small image area and are easily confused with complex background structures. To address unreliable ranking of weak foreground candidates and insufficient local-detail retrieval in Detection Transformer (DETR)-based detectors, we propose the Prior-Driven and Detail-Refined Detection Transformer (PDR-DETR), built on the Real-Time Detection Transformer (RT-DETR). PDR-DETR introduces two stage-specific modules. Query-guided frequency prior injection (QFPI) converts target-related spectral responses into priors for encoder-memory recalibration and multi-scale query ranking, improving the selection of weak foreground candidates. Reference-guided cross-scale local aggregation (RCLA) retrieves candidate-centered local evidence from high-resolution features to refine intermediate decoder queries. Experiments on a self-collected five-category railway signal light dataset containing 12,810 images show that PDR-DETR achieves 0.580 AP50:95 and 0.561 APs, improving the RT-DETR baseline by 0.041 and 0.037, respectively. These results demonstrate the effectiveness of the proposed prior-to-detail detection framework for small railway signal light detection in complex railway environments.
Funding Information
  • National Natural Science Foundation of China (52472422, 52502453, 52477065, and 52293424)

This publication has 28 references indexed in Scilit: