Validating traffic models using large-scale automatic number plate recognition (ANPR) data

Abstract

Traditional manual survey methods for collecting reliable origin-destination data to develop large strategic transport model is notoriously expensive and the sample sizes are often relatively small. Arguably, the least reliable data required for the development of strategic traffic models is the origin-destination data. Recent technological advances, such as probe data from on-board devices, have been successful in providing data for some needs such as journey times and routing options. However, varying degrees of success have been achieved in obtaining reliable origin-destination (OD) data from these new technologies. Automatic Number Plate Recognition (ANPR) is one if the newer technologies that could be used to collect large-scale data sets over the large study areas that strategic traffic models cover. The aim of this study is to examine ANPR data collected from the Gauteng Freeway Improvement Project's (GFIP) Open Road Tolling (ORT) gantries in terms of its accuracy and uses in the development and improvement of strategic traffic models. Of particular interest is the use of the ANPR data to contribute towards the improvement of the distribution of trips in the OD matrices. This is achieved by developing methodologies to derive comparable gantry to gantry traffic volumes from the ANPR data and the GFIP traffic model. The above comparisons enabled the undertaking of a post opening project evaluation of the GFIP traffic model's 2015 forecasts using as many characteristics of the traffic flows and patterns that can be derived from the ANPR data. Characteristics such as traffic volumes and journey times are directly comparable with standard traffic model outputs. Tracking vehicles between gantries enabled the calculation of the number of trips that travel between gantry pairs giving rise to gantry-to-gantry (G2G) trips, which can be represented in a G2G count matrix. This G2G count matrix has probably the most beneficial data that can be derived from the ANPR systems as it contains an "accurate" element of the trip distribution on the road network. A methodology was developed to derive equivalent trip matrices from a traffic model's select-link trip matrices where the links are those where the gantry (ANPR camera) is located. The sums of the trips in the derived sub-matrices match the G2G counts. This enabled the comparison between the modelled trip distribution represented by the select link to select link (SL2SL) volumes and the actual ANPR G2G counts. This is in fact a comparison of a portion of the model's distribution to actual, comprehensive data. This study demonstrates that ANPR data has the potential to improve strategic traffic models. The automation of the processes to derive the SL2SL assigned volumes from the models and combining it with existing matrix estimation techniques will enhance the trip distribution in the output trip matrix. The current practice of using individual traffic counts in matrix estimation has the adverse tendency to affect the trip distribution. Hence, the recommendation to use traffic counts in matrix estimation to traffic counts with caution.