**6. Conclusions**

In this research, the cyclists' flows obtained by traditional counting methods have been compared with GPS traces from smartphones at the same locations and during the same time period.

Although crowdsourced cyclists represent often a small portion of all cyclists, they do represent well the ridership of Bologna in terms of cyclists' volumes and gender distribution. This result emerges clearly by comparing traditional counting method with GPS traces, confirming their representativeness of the population. The correlation between cycling counts and GPS data collected by smartphones has been relatively high, with an R<sup>2</sup> value of 0.73. This correlation is significantly higher than the results obtained by other studies, most likely due to the more detailed representation of the Bologna network, including footpaths in parks and the possibility to cycle one-way roads in both directions. Due to this high correlation, it has been possible to estimate the absolute bicycle flows on all network links by an appropriate scaling of the map-matched flows. The cyclists' routes are of grea<sup>t</sup> value for the planning of cycling infrastructure and the drafting of cycling policies. The proposed method, which combines bicycle counts at a few main road sections with areas covering GPS traces, can readily be applied in other cities in order to reliably estimate the absolute bike flows of an entire urban area.

GPS data have been further used to determine the total deviation metric, which counts the total deviations that a road link causes to cyclists. The total deviation metric is useful to identify weak links of the cycling network, but it does not identify the reason why certain road links are avoided. However, applying the total deviation metric to the Bologna road network, the highest deviation has been seen on tra fficked roads without physically protected bike lanes. Also, roads with reserved bus lanes, which are open for bicycles too, showed high deviation rates. Further analyses of chosen and shortest road sections have shown that cyclists are willing to make deviations when the alternative route provides a high share of reserved bikeways, a high share of low-priority lanes, a low intersection density and a low share of roads with mixed tra ffic (cyclists with buses and pedestrians). Planners should take the deviation metric into considerations for either bike-path construction or bike-network interventions. Obviously, the total deviation metric does not reveal deviations if there are no alternatives to avoid a certain road link. The map-matched traces allowed to calibrate a discrete choice model between two route alternatives, considering distance, share of exclusive bikeway and share of low-priority roads. A longer distance and a higher share of low-priority roads appear to decrease the choice probability, while

a higher share of exclusive bikeways does increase the choice probability, as expected. With the same model, it has been possible to quantify the tolerated deviation length in function of the road attributes.

In future works, the representativeness of the results could be improved by statistically weighting the GPS traces according to different person attributes, such as occupation, gender, or age. The route choice model could be enriched by more significant attributes like traffic light density, junctions with left-turns, or junctions with side-roads entering from the right side. In particular, the low-priority road attribute needs to be further refined. The generation of longer route alternatives could be replaced by actually chosen routes and models using only attributes of the chosen routes shall be tested.

**Author Contributions:** Study conceptualization, Federico Rupi and Joerg Schweizer; Methodology, Federico Rupi and Joerg Schweizer; Programming and data retrieval, Federico Rupi, Joerg Schweizer and Cristian Poliziani; Formal GIS analysis, Joerg Schweizer and Cristian Poliziani; Writing—review and editing, Federico Rupi, Joerg Schweizer and Cristian Poliziani.

**Funding:** Open access funding provided by Department of Civil, Chemical, Environmental, and Materials Engineering (DICAM)—University of Bologna.

**Acknowledgments:** We are grateful to SRM Bologna srl for providing the GPS data of the European Cycling Challenge 2016.

**Conflicts of Interest:** The authors declare no conflict of interest.
