Robustness of automated motorcycle helmet use detection in real-world application

Context
The motorcycle helmet is the most effective injury prevention measure in case of a crash. Still, many countries do not have adequate motorcycle helmet use data, preventing targeted legislative or enforcement-based interventions. A primary obstacle to assess the helmet use of riders is the use of human observers, which are expensive to employ and only register motorcycle helmet use for small amounts of traffic. Previous research has focused on automated helmet use registration methods for individual countries. However, this country-specific approach raises the question of whether methods can be easily transferred to different countries with new roads while still producing accurate helmet use predictions.

Methods
A computer vision-based algorithm for detecting motorcycle helmet use was trained and tested on Myanmar road traffic videos to register helmet use rates. The Myanmar-trained machine learning model was then transferred onto video data from Nepal to register helmet use there.

Results
The computer vision algorithm detected 67.3% of all active motorcycles in the Nepal video data. Human-registered helmet use in the Nepal data was 73.4%. In comparison, the algorithm-registered helmet use rate was 57.2%, i.e. there was an underestimation of helmet use by 16% in the automated helmet use detection.

Conclusion and Learning Outcomes
Computer vision-based helmet use detection algorithms are highly accurate in the road environments they were trained on. This study shows that the application on new road environments leads to a decrease in accuracy. Hence more research into algorithm-based motorcycle helmet use detection is needed to increase robustness.