A two-stage cascade framework for traffic sign detection: reproduction and boundary analysis

Traffic sign detection in adverse conditions remains difficult because small, distant signs and degraded illumination act together. This paper reports a reduced-scale reproduction and boundary analysis of a cascade framework for traffic sign detection, comprising multi-dimensional preprocessing, a fast-screening / fine-recheck cascade, and a rule-based semantic gate. We characterise the framework by condition, showing where it works and where it fails. On TT100K (45 classes; 3016-image official test split), a YOLOv8n screening stage trained for 20 epochs reaches mAP@0.5 of 0.187 ± 0.005 over three seeds. Component ablations show that the full pipeline does not outperform a confidence-only cascade on clean data, and that preprocessing yields no measurable benefit in clean conditions. Cross-dataset evaluation on CCTSDB reveals monotonic degradation with sign size (recall 0.323 for XL, 0.002 for XS) and pronounced weakness at night and in rain; preprocessing improves F1 by 30% at night but is ineffective in rain and for small signs. A rule-based semantic gate modifies only 0.66% of detections and loses true positives in 29% of its changes. End-to-end latency on our hardware is 239 ms (median), dominated by CPU preprocessing, giving a single-stream capacity of 1.6 frames/s. We conclude that the cascade skeleton is implementable and that small-sign detection and preprocessing cost are the binding constraints.