PT  - JOURNAL ARTICLE
AU  - Hüttel, Frederik Boe
AU  - Riis, Christoffer
AU  - Kristensen, Jakob
AU  - Gül, Oguzhan
AU  - Janstrup, Kira Hyldekær
AU  - Lin, Hanhe
AU  - Pereira, Francisco Camara
AU  - Siebert, Felix Wilhelm
TI  - 463 Using computer vision for the automated detection of cyclists’ safety-related behaviour
AID  - 10.1136/injuryprev-2022-safety2022.210
DP  - 2022 Nov 01
TA  - Injury Prevention
PG  - A71--A71
VI  - 28
IP  - Suppl 2
4099  - http://injuryprevention.bmj.com/content/28/Suppl_2/A71.1.short
4100  - http://injuryprevention.bmj.com/content/28/Suppl_2/A71.1.full
SO  - Inj Prev2022 Nov 01; 28
AB  - Context The COVID-19 pandemic has increased bicycle use in cities worldwide as citizens shift from public transport towards individual transportation modes. At the same time, cycling infrastructure is slow to adapt to this increase in traffic. For crowded infrastructure, the safe behaviour of cyclists increases in importance. But while automated solutions for detecting rule infractions of motorised vehicles exist (e.g. speeding-cameras for cars), no such systems for detecting dangerous behaviours for cyclists are available, preventing road safety actors from conducting focused education campaigns.Methods We collected roadside video data containing 645 cyclists on multiple cycling lanes in Copenhagen, Denmark. Human observers pre-screened the video data and identified instances of unsafe behaviour, i.e. cyclists not using a helmet, using a phone, or using headphones. We trained a state-of-the-art computer vision algorithm to detect these unsafe behaviours in cyclists. The algorithm was then applied to test data, which contained 126 cyclists in total.Results Overall, the algorithm detected 89.5% of the bicycles present in the test video data, i.e. 15 out of 126 cyclists were not detected. Unsafe behaviour of cyclists was correctly identified in 77% of instances (missing 18 occurrences). Generally, non-helmet-use detection worked best, while performance for headphone-use detection was the lowest.Conclusion and Learning Outcomes There is a lack of safety-related behavioural data for bicyclists. The automated detection of safety-related behaviour through computer vision algorithms presents an efficient way to collect comprehensive, diverse and temporal data, which policymakers can use to plan police enforcement and traffic education campaigns.