Reducing Cyclist Crashes by Assessing the Road Environment: An Application of Google Imagery and Machine Learning

Abstract

Background: Cycling is an active and sustainable transportation mode, and is associated with health, environmental and societal benefits. Therefore, increasing the use of bicycles is being supported as a transport policy in many countries. However, despite these benefits, cyclists are vulnerable road users and are over-represented in traffic crash casualties compared to other modes of transport. The injury concern can discourage people from adopting cycling as a main transportation mode. Urban infrastructure that caters to cyclists' safety can potentially reduce crashes and therefore, injury morbidity and mortality. Methods: This research uses cyclist crashes recorded by the state road authority from 2010 to 2013 in Greater Melbourne. Exposure data used anonymised bicycle trips recorded by volunteer users of RiderLog smartphone application from 2010 to 2013. Crash locations and control sites were sampled from areas with high cycling exposure. Google Street View maps and satellite images at crash locations and control sites were downloaded to capture information of the road environments where cyclists crash and never crash. Deep learning methods using generative adversarial networks were applied to explore features of road environments associated with cyclist crashes. Results: A number of unique observations were identified namely, that locations that have low crash risk had more green space (trees or grass), and median strips (that separate traffic from opposing lanes on divided roadways) also decreased a cyclist’s crash risk. Road environments with high-rise buildings casting shadows on the roadside are mostly seen in the environment in which crashes occurred. The experiments also identified factors that have been reported previously in the literature and statistical analysis, providing confidence in the presented methods. Such factors include tram tracks, intersections, on-road parking and off-road bicycle paths. Statistical analysis showed 52.6% of crash locations were within 5 metres of a tram line, while this percentage for control sites was 5.6%. Conclusions: This research presents a method that takes advantage of the increasing availability of big datasets, computing power and the advances of deep learning techniques, to analyse the road environments of locations where cyclists crash from a new perspective. The findings give urban planners insights on how streetscapes might be reconstructed to improve safety situations for cyclists. The results also provide transportation engineers and cyclists with visual indications about what kind of streetscapes are safer.