*PhD Defense*
December 1 2014 09:30–11:30
4080 AIR Building
Sarah Hernandez
Sarah Hernandez
TSE PhD Candidate

Transportation agencies tasked with forecasting freight movements, creating and evaluating policy to mitigate transportation impacts on infrastructure and air quality, and furnishing the data necessary for performance driven investment depend on quality, detailed, and ubiquitous vehicle data. Unfortunately, commercial vehicle data is either missing or expensive to obtain from current resources. To overcome the drawbacks of existing commercial vehicle data collection tools and leverage the already heavy investments into existing sensor systems, we present a novel approach of integrating two existing data collection devices to gather high resolution truck data – Weigh-in-motion (WIM) systems and advanced inductive loop detectors (ILD). Each source provides a unique data set that when combined produces a synergistic data source that is particularly useful for truck body class modeling. Since body configuration is closely linked to commodity carried, drive and duty cycle, and other operating characteristics, it is inherently useful for each of the above mentioned applications.

In this work we describe the physical integration including hardware and data collection procedures undertaken to develop a series of truck body class models. Approximately 33,000 samples consisting of photo, WIM, and ILD signature data were collected and processed representing a significant achievement over previous ILD signature models which were limited to around 1,000 commercial vehicle records.

Three families of models were developed, each depicting an increasing level of input data and output class resolution. The first uses WIM data to estimate body class volumes of five semi-trailer body types and individual predictions of two tractor body classes for vehicles with five axle tractor trailer configurations. The trailer model produces volume errors of less than 10% while the tractor model resulted in a correct classification rate (CCR) of 92.7%. The second model uses ILD signatures to predict 47 vehicle body classes using a multiple classifier system (MCS) approach coupled with the Synthetic Minority Oversampling Technique (SMOTE) for preprocessing the training data samples. Tests show the model achieved CCR higher than 70% for 34 of the body classes. The third and most complex model combines WIM and ILD signatures using to produce 63 body class designations, 52 with CCR greater than 70%. To highlight the contributions of this work, several applications using body class data derived from the third model are presented including a time of day analysis, average payload estimation, and gross vehicle weight distribution estimation.