Answers are grounded in this paper and cite their sources.
Automated helmet monitoring can miss visible helmets when supplementary sun-protective headwear (SPH) alters appearance. HELMET evaluates five YOLO detectors under native input and bilinear/ESPCN upscaling (×2/×3) using 1,340 Thai construction-site images with 3,663 annotated heads. Testing uses 134 held-out images containing 378 heads, including 296 visibly helmeted heads. False non-compliance rate (FNCR) measures failure to credit helmet presence, not regulatory compliance. Native FNCR is 0.375 for the original generic detector and 0.135 for the binary contextual detector, although architectures and class spaces differ. The generic detector shows no significant improvement across four operating-point metrics (adjusted p ≥ 0.12); contextual F1 declines from 0.868 to 0.711–0.820. ESPCN ×2 reduces FNCR in the four-class merged model from 0.270 to 0.199 (adjusted p = 0.022 for paired recall). Findings support empirical preprocessing assessment, limited by one site and fixed checkpoints. Helmet recognition cannot establish protective adequacy.
No takes yet. Share an insight, caveat, or question.
Malaikrisanachalee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: