Cycling near misses: a review of the current methods, challenges and the potential of an AI-embedded system

Whether for commuting or leisure, cycling is a growing transport mode in many countries. However, cycling is still perceived by many as a dangerous activity. Because the mode share of cycling tends to be low, serious incidents related to cycling are rare. Nevertheless, the fear of getting hit or falling while cycling hinders its expansion as a transport mode and it has been shown that focusing on killed and seriously injured casualties alone only touches the tip of the iceberg. Compared with reported incidents, there are many more incidents in which the person on the bike was destabilised or needed to take action to avoid a crash; so-called near misses. Because of their frequency, data related to near misses can provide much more information about the risk factors associated with cycling. The quality and coverage of this information depends on the method of data collection; from survey data to video data, and processing; from manual to automated. There remains a gap in our understanding of how best to identify and predict near misses and draw statistically significant conclusions, which may lead to better intervention measures and the creation of a safer environment for people on bikes. In this paper, we review the literature on cycling near misses, focusing on the data collection methods adopted, the scope and the risk factors identified. In doing so, we demonstrate that, while many near misses are a result of a combination of different factors that may or may not be transport-related, the current approach of tackling these factors may not be adequate for understanding the interconnections between all risk factors. To address this limitation, we highlight the potential of extracting data using a unified input (images/videos) relying on computer vision methods to automatically extract the wide spectrum of near miss risk factors, in addition to detecting the types of events associated with near misses.