Evidence-Aware Adaptive Recognition for Intelligent Traffic Sign Systems with Calibrated Uncertainty under Distribution Shift

Abstract Traffic-sign recognition under image shiftraises two related questions: which label is supportedand whether the crop still contains the visual evidenceon which that decision is based. Adaptive signconditionedrecognition (ASCR) keeps these quantitiesseparate. A temperature-calibrated classifier withregularized adaptive prediction sets (RAPS) returnssemantic probabilities and label sets; a label-free monitormeasures preserved color, area, shape, and componentstructure. Track-disjoint traffic-sign data andcontrolled corruptions reveal a high-confidence/lowevidenceregime: the lowest-evidence group retains 0.913confidence at 0.162 error. The evidence–error slopealso changes sign across the evaluated environments,indicating that a single global evidence coefficient canbe misleading under these shifts. ASCR therefore actsat the decision stage: semantic ambiguity expands thelabel set, while evidence departure triggers intervention.Within the evaluated settings, this separation providesa transparent decision interface for calibrated classifiersand visual-evidence monitors.