Estimating city-wide hourly bicycle flow using a hybrid LSTM MDN

Cycling can reduce greenhouse gas emissions and air pollution and increase public health. With this in mind, policy-makers in cities worldwide seek to improve the bicycle mode-share. However, they often struggle against the fear and the perceived riskiness of cycling. Efforts to increase the bicycle's mode-share involve many measures, one of them being the improvement of cycling safety. This requires the analysis of the factors surrounding accidents and the outcome. However, meaningful analysis of cycling safety requires accurate bicycle flow data that is generally sparse or not even available at a segment level. Therefore, safety engineers often rely on aggregated variables or calibration factors that fail to account for variations in the cycling traffic caused by external factors. This paper fills this gap by presenting a Deep Learning based approach, the Long Short-Term Memory Mixture Density Network (LSTMMDN), to estimate hourly bicycle flow in Copenhagen, conditional on weather, temporal and road conditions at the segment level. This method addresses the shortcomings in the calibration factor method and results in 66-77% more accurate bicycle traffic estimates. To quantify the impact of more accurate bicycle traffic estimates in cycling safety analysis, we estimate bicycle crash risk models to evaluate bicycle crashes in Copenhagen. The models are identical except for the exposure variables being used. One model is estimated using the LSTMMDN estimates, one using the calibration-based estimates, and one using yearly mean traffic estimates. The results show that investing in more advanced methods for obtaining bicycle volume estimates can benefit the quality, mitigating efforts by improving safety analyses and other performance measures.