Automated Validation and Interpolation of Long-Duration Bicycle Counting Data

Abstract

Bicycle flow data is crucial for transportation agencies to evaluate and improve cycling infrastructure. Average annual daily bicyclists (AADB) is commonly used in research and practice as a metric for cycling studies such as ridership analysis, infrastructure planning, and injury risk. AADB is estimated by averaging the daily cyclist totals measured throughout the year using a long-term automated bicycle counter, or by using long-term bicycle counting data to extrapolate data from a short-term counting site. Extrapolation of a short-term bicycle counting site requires an accurate and complete set of daily factors from a group of references: long-term bicycle counters. In practice, validation of reference data is done manually, an exercise that is time-consuming but crucial as significant error can be introduced into AADB extrapolation if reference data are not validated. This paper proposes an automated method to validate long-term bicycle count data and interpolate anomalous portions of data. As part of this work, the methods are validated using a relatively large dataset of automated bicycle counts. For validation of our approach, data anomalies are created artificially in a way that removes data (first trial), or reduces counts to 25% or 40% of the measured bicycle counts (second and third trials), for 6 hours, 12 hours, and full days. Of the more than 100 generated anomalies, the validation process flagged approximately 90% in the first and second trials and 80% in the third trial. The average absolute relative error of the interpolated daily values was approximately 10% for all three trials.

Daily Traffic Count Imputation for Bicycle and Pedestrian Traffic: Comparing Existing Methods with Machine Learning Approaches

“…Table 2 shows the mean APE (MAPE) and median APE for annual traffic estimation using each of the tested methods. These results were split into two groups according to whether the sites could be matched to a factor group based on the criterion recommended in Beitel et al (22) and El Esawey (23). This process required that the daily counts were correlated with the reference group at the 0.75 level and that at least three sites were included in the reference group.…”

“…The factors from the matched site are then applied to the counts available at Site j , and then aggregated for an estimated annual and average annual traffic count. Beitel et al recommend matching sites using a correlation coefficient of 0.75, which this research followed (22). In addition, these recommendations suggest making sure that at least three sites are present in the group of matched reference sites, which is not always attained in this deployment of the DOY method.…”

“…The section describes the different methods tested to determine performance of various imputation approaches, as well as detailing the experimental method deployed to test the accuracy of each method. The first method used a DOY factoring approach following the methods documented in Beitel et al, where Site i is matched to one or more sites using the measured correlation of their DOY factors (22). DOY factors are calculated by dividing the daily traffic count by the annual traffic total.…”

“…This approach however, requires enough data and counters to match the traffic count site to a reference site or group with similar DOY factors, which is not always possible. If conditions can be met for this method, error appears to be low, though in this research Beitel et al did not examine the ability of the method to impute pedestrian counts (22). Since the DOY method works best when used to expand short-duration counts and has been used in research investigating a missing data problem, this method served as a baseline for comparing the methods proposed in this research, with more discussion below.…”

“…This process introduces unknown error when applying factors if short-duration sites are not properly matched to a reference site or group or if insufficient sites exist in the expansion group (15). Beitel et al enlists the DOY method for a data imputation problem to test the accuracy of imputation using DOY expansion factors (22). The authors propose a method for matching sites based on a correlation coefficient of 0.75 for sites where the annual average daily bicycle traffic is 250 bicycles per day or more.…”