A METRANS funded research paper authored by USC researchers Dr. Cyrus Shahabi, Dr. Luciano Nocera, and Dr. Genevieve Giuliano studied the feasibility of freight traffic modeling from precise and localized but sparse freight data and built machine learning (ML) models that can predict freight volumes on highway segments. Freight data sources are very limited, often based on infrequent surveys or extrapolations of in-road sensor data. This research aims to close that gap by combining multiple data sources - e.g., weigh in motion stations (WIMS), the UCI Truck Activity Monitoring System (TAMS), and CCTV cameras which are now widely available on highways. The contribution of this research is to use advanced ML models to process data (e.g., from CCTV video feeds) to extract truck observations and hence more accurate truck flow data. 

The research team focused on a restricted area in the Los Angeles metropolitan area where freight volume is most relevant. They developed a research strategy that proceeds with limited data by (1) generating synthetic datasets on the area of interest that can be used to train and verify different freight volume estimation models, and (2) leveraging existing Caltrans CCTV video footage from public Caltrans feeds to investigate whether cameras can be used to provide truck observations (i.e., timestamp, truck type.)

Their simulation software allowed them to generate consistent datasets of truck traffic over a highway system, including the number and type of trucks, their departure and destination, the number of sensors, and truck observations at sensor locations. The simulation software produces truck trajectories and truck observations that are consistent with traffic conditions that we obtain from traffic data collected by the USC METRANS Archived Data Management System (ADMS), which archives over ten years of freeway and arterial sensor data. This allowed the research team to generate truck traffic patterns under different traffic conditions and fleet mixes. The figure below shows inferred truck volume reconstructed from synthetic data for different approaches. The results indicate that ML methods can increase the accuracy of truck counting and hence truck origin-destination matrices, improving traffic modeling.

One of the biggest challenges in freight transportation planning is the lack of data on truck movements. For effective freight planning we need information on the origins, destinations, and routes of truck trips. GPS data is the best available source, but most GPS data are private and unavailable to government agencies. This research offers a privacy-preserving, cost-effective approach for estimating truck traffic using existing roadside infrastructure, such as CCTV cameras and weigh-in-motion sensors. By enabling accurate freight movement analysis without relying on private GPS data, the findings can support more informed transportation planning, infrastructure investment, and freight policy decisions.

Read the full report here: https://www.metrans.org/research/validation-of-freight-volume-modeling-on-major-highway-links.