Enhancing Insights into Traffic Flows and Activities: Evaluating and Exploiting Machine Learning Algorithms in Real-World Scenarios

Understanding truck activities has become increasingly crucial in traffic research, considering the increase in electric vehicles, the potential failure of critical infrastructure, etc. There are many different data sources to monitor the traffic flow. In this study, four different data sets generated from different approaches are used to extract traffic information. An innovative approach is devised and implemented for each data set to get valuable insights. Chapter I improves a recent Linear Programming method to tackle the truck identification problem based on the results of the radar detector or its equivalents (e.g., single loop detector). Tested under different contexts, the method is simple yet stable. It outperforms the two classical methods in many cases. After discussing identifying trucks for the most equipped radar detectors, Chapter II focuses on extracting traffic information from CCTV cameras, another popular traffic detector. The problem is challenging because CCTV video’s resolution can be very low. To solve the problem, a framework with four modules is proposed to tackle four different subtasks, including vehicle detection, classification, tracking, and information aggregation. Chapter III tries to impute traffic flow by vehicle types over the national highway network. A new method is proposed by interactively imputing traffic information over the network and generating traffic volume data for each vehicle class under the prior assumption of Annual Average Daily Traffic (AADT) flow is known. Chapter IV uses survey data to understand the shipper’s decision. The survey data, namely the Commodity Flow Survey (CFS), is collected from real freight choice of shippers. Besides applying machine learning methods to predict the freight mode choice, ensemble techniques are used to achieve better prediction performance. SHAP values are computed to interpret the machine learning model. Four chapters tackle the problem of understanding truck traffic from different perspectives. The first two chapters focus on improving local estimates, which mainly benefit the truck estimated at corridor or metropolitan levels. The last two chapters generate the whole picture of truck flows at a state or national level, a critical ability for infrastructure planning, sustainability analysis, emission modeling, etc.