Bikeable: Estimating cyclists’ perceived safety from street-level imagery to support transport planning in Lisbon and London

There has been a steady increase in the number of people using micromobility as a mode of transportation. Micromobility offers several advantages over other forms of transport, from improved physical and mental health for users, to reduced noise and air pollution. However, the downside of the rise in micromobility usage is that it has also led to an increase in accidents involving these vehicles. Many of these accidents result in severe injuries or even fatalities. Therefore, people often choose not to use micromobility due to safety concerns, despite the numerous benefits. To tackle this issue, we aim to help individuals avoid dangerous roads and high-risk exposure by using artificial intelligence. To achieve this goal, a Multilayer Perceptron neural network called Bikeable has been trained to predict a safety score from an image. The training input for the neural network included the collection of 100 thousand perceived safety scores, as well as information on objects detected through Object Detection and structures identified through Image Segmentation. The final accuracy achieved was 82.19 ± 2.44% after being 5-fold cross-validated and 79% in the final test evaluation. This article sets the ground for an accurate and scalable calculation of micromobility riders’ perception of safety and the future development of safety-centered route planners.