One-Dimensional Convolutional Neural Network Model for Local Road Annual Average Daily Traffic Estimation

Abstract

The focus of this research is to develop a robust model for accurately estimating link-level annual average daily traffic (AADT) of all the local functionally classified roads. The capabilities of one-dimensional convolutional neural network (1D-CNN), a deep learning architecture, and the domain knowledge pertaining to local road travel characteristics were combined to estimate local road AADT. The AADT based on traffic counts collected at 12,769 traffic count stations on local roads in North Carolina during 2014, 2015, and 2016 were considered for model training, validation, and testing. A total of eight existing state-of-the-art statistical, geospatial, and selected other machine learning models were compared with the 1D-CNN model to estimate local road AADT. These include ordinary least square (OLS) regression, geographically weighted regression (GWR), ordinary kriging, natural neighbor (NN) interpolation, inverse distance weighting (IDW), backpropagation artificial neural network (BP-ANN), random forest (RF), and support vector machine (SVM). The model development and test results showed that the 1D-CNN model performed better than the other considered models. The architecture of the 1D-CNN model can learn the intricate patterns in the local road AADT. The outputs from the methodological framework proposed in this research help practitioners perform safety evaluation, planning and implementing infrastructure improvements, fund allocation and prioritization, air quality estimates, and meeting Highway Safety Improvement Program (HSIP) reporting requirements.

Keywords

• Annual average daily traffic
• AADT
• convolutional neural networks
• deep learning
• local road
• low-volume road