Estimation of daily bicycle traffic volumes using sparse data

Abstract

Calculating the annual average daily bicycle (AADB) volume at a particular cycling facility requires the availability of year-round daily volume data. Automatic counters (e.g., loop detectors) used to collect such continuous data are subject to periodic malfunctions, leading to sporadic data gaps. This problem could affect the calculated values of the AADBs and impact the estimates of the daily and monthly adjustment factors at these count stations. The impacts become even more significant if the data gaps take place frequently and/or for long periods. This research tackles the problem of missing cycling traffic volumes at count locations that potentially experience frequent sensor malfunctions during the year. The method is also applicable to any other similar research problem (e.g. missing volumes at vehicle count locations). A data-driven, yet novel, model is proposed to estimate missing volumes at some locations, using data from other nearby count locations as well as the historical volumes of the same location. The model is motivated by the spatial-temporal relationship of cycling volumes of similar nearby facilities. The proposed model is dynamic as it assumes no prior knowledge about which locations may experience sensor malfunction (i.e., missing volumes). The model is referred to as the "autoencoder neural network" and it belongs to the family of Artificial Neural Networks (ANNs). This model expresses the relationship between a vector of input variables and itself. Hence, if a daily bicycle count is available on one day at a specific location, it will be used as a model input; whereas, a missing daily volume will be treated as an output variable that needs to be determined. The model was tested using a large dataset of about 13,000 daily bicycle volumes from the City of Vancouver, Canada. The data were collected between 2009 and 2011 at 22 different count locations. The model showed a strong estimation power with an average error of about 10%. Sensitivity analyses were carried out to investigate the impact of different model parameters on the estimation accuracy. The optimum set of model parameters was consequently defined.