Background Motorcycle crashes carry considerable head injury risks. Hence, helmets are crucial for reducing the severity of injuries and fatalities. Despite mandatory helmet use laws in many countries, the lack of comprehensive and timely data on helmet use impedes the identification of urban areas with low helmet use, preventing targeted legislation and educational campaigns.

Objective This study employs a cost-efficient computer vision methodology using crowdsourced images to gather data on city-specific motorcycle helmet use from five Southeast Asian capital cities.

Methods We utilized Mapillary, a crowdsourced image platform, to gather street-level image data recorded between 2015 and 2021 from Bangkok, Manila, Kuala Lumpur, Jakarta, and Hanoi. We employed a state-of-the-art object detection algorithm trained on more than 800,000 images to detect motorcycles, rider numbers, and helmet use of drivers and passengers.

Results On an annotated test dataset, the algorithm achieves high accuracy for helmet use detection, with algorithm-registered helmet use differing by only by 1.9% from human-registered helmet use. Applying the algorithm to more than 600,000 images from the five cities detects more than 1.3 million motorcycles in the dataset. Temporal fluctuations in the observed urban environments are identified, with average yearly fluctuations of urban motorcycle helmet use within the five observed cities by 19.1%. The automated helmet use registration further reveals known patterns of rider-position and observation location-specific helmet use. In one example from Bangkok for rider position, helmet use of drivers was registered to be considerably higher than passenger helmet use (81.7% vs. 70.9%). For observation location, helmet use on primary roads is found to be higher than on secondary roads for all investigated cities.

Conclusions Our study presents a cost-efficient method for generating accurate, city-specific motorcycle helmet use data in Southeast Asia using computer vision on diverse crowdsourced imagery. This enables the registration of temporal and location-specific motorcycle helmet use patterns without costly on-the-ground data collection. Further potential for scaling the proposed approach, as well as methods for validation of the registered data, are discussed.