MotoHalo: Automated Helmet Compliance Monitoring Using YOLO-Based Localization and Deep Ensemble Classification

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Preventable road-traffic injuries remain a serious concern where motorcycle-helmet regulations are not consistently enforced. Monitoring is particularly difficult in urban traffic, where helmet evidence may be small, partially occluded, or blurred within cluttered rider-vehicle scenes. The present work develops a framework for monitoring helmet compliance that is comprised of two decoupled stages. The first stage pertains to the monitoring region of interest for traffic, and the second stage provides classification of a scene into one of four categories— Helmet, No_Helmet, Phone, or Rider. The proposed system provides a filtered output that is of primary interest to enforcement, in the form of a binary visualization of Helmet or No_Helmet. Within the proposed work, multiple backbone convolutional networks were explored to achieve the desired goals. ConvNeXt-Tiny and ConvNeXt-Small provided the best results and achieved a single-model validation accuracy of 96.72%. The highest single-model macro F1 score was achieved by ConvNeXt-Small with a score of 0.92. A soft voting ensemble of ConvNeXt-Tiny, ConvNeXt-Small and ResNet-50 improved validation accuracy on the original dataset to 97.16%. An equal-weighted ensemble of the same dataset validated at 99.00% and tested at 98.22%. On the designated frozen MakeML test set, MotoHalo achieved 96.98% balanced accuracy, compared with 89.31% for the single-stage YOLO baseline, showing that the proposed decoupled detection-classification approach can improve helmet-compliance assessment under the evaluated conditions.