Abstract
Vehicles are classified as 13 classes in the Federal Highway Administration vehicle classification scheme, including motorcycles, passenger cars, buses, and various types of trucks. Truck traffic exerts a significant impact on pavement structure and is the main force causing pavement structure damages and distresses. Detailed truck traffic data are continuously recorded using Weight-in-Motion devices on the highway system. There usually exist a noticeable amount of unclassified and misclassified vehicles in the recorded dataset, which are those that the devices failed to identify their vehicle types based on the integrated criteria. As truck traffic is the main input of mechanistic-empirical pavement design, inaccurate vehicle classification would cause overdesign or underdesign of pavement structures and would consequently result in unnecessary extra construction costs as well as highway user costs. This paper presents a method to identify and reclassify the misclassified and unclassified vehicles utilizing the convolutional neural network in combination with spatial pyramid pooling. Through fine-tuning a series of hyperparameters, the proposed model successfully identified and reclassified a major portion of the unclassified and misclassified vehicles.
