|
Abstract
|
Proceedings of the 14th World Conference on Injury Prevention and Safety Promotion (Safety 2022)
Context Motorcycle helmet use represents one of the main progress metrics in the Second Decade of Action for Road Safety and the Sustainable Development Goals. Still, data on this critical variable is only infrequently collected, hence no evidence based and timely interventions to influence the helmet use of motorcyclists are possible. A main obstacle is the complex data collection process, involving either human observers, or the installation of dedicated traffic observation cameras. However, crowd-sourced data from e.g. Google Streetview or Mapillary is readily available for processing, and automated helmet use detection on those images would allow efficient global data collection.
Methods More than 800,000 on-road images from Bangkok (Thailand) and Jakarta (Indonesia) were assessed on the Mapillary platform. A state-of-the-art object detection algorithm was used to identify images that include motorcycles. In a subset of the identified images, motorcycle rider numbers and helmet use were annotated. The algorithm was then retrained to detect active motorcycles, rider numbers, and helmet use, and tested on 600 new images from the Bangkok and Jakarta road environment, which had not been part of the training.
Results General motorcycle detections by the trained algorithm were correct in 91% of cases. Helmet use in the test data was underestimated by approximately 15% (human-observed helmet use: 67.8%; algorithm detected: 53.1%).
Conclusion and Learning Outcomes The application of computer vision to crowd-sourced images can generate motorcycle helmet use data for a fraction of the costs of traditional data collection, but accuracy needs to be improved further.
Language: en |