Source: samafoto via iStock
One of the most challenging problems in urban transportation planning is the lack of fine-grain data on freight movements. Cities and regions do not know how many trucks operate on their streets and highways and have limited freight flow information. Recent PSR- funded research shows the potential for an effective and scalable solution to truck flow estimation using advanced computational methods applied to data from existing sensors.
In the recently published research report Validation of Freight Volume Modeling on Major Highway Links, METRANS faculty researchers Dr. Cyrus Shahabi, Dr Luciano Nocera, and team develop methods to generate freight volume information such as hourly origin-destination count (OD-matrices) directly from sensor observations. The research team studied truck traffic on highways in an extended region in Los Angeles County, approximately 12 square miles in size, which includes major truck corridors located to the north and east of the Ports of Los Angeles and Long Beach, which combined are the busiest container gateway in North America. The researchers focused on using sensors currently deployed along roadways, including close circuit television (CCTV) cameras, Weight-In-Motion Stations (WIM) sensors, and other sensors such as Truck Activity Monitoring Systems (TAMS).
With the help of Caltrans District 7, the California Department of Transportation Traffic Operations Division, the UC Irvine Institute of Transportation Studies, and the Integrated Media Systems Center at USC, Shahabi and Nocera collected CCTV videos, Weigh-in-Motion (WIM) station data, and TAMS data. Additionally, they used data from their almost twelve-year archive of Los Angeles Metro highway sensor data, the Archived Data Management System (ADMS). The resulting dataset was used to study how well one can detect and count trucks and integrate these observations over the road network to model trucks' flow.
Deep learning-based object detection and tracking algorithms were applied to Caltrans CCTV videos to study the feasibility and usefulness of using traffic monitoring videos for truck detection and counting. The team considered four classes of trucks and evaluated the model’s performance by comparing the results with ground truth counts. They showed that truck detection and counting can be achieved with high detection accuracy and low inference times, allowing them to deploy these methods on video streams. Additionally, robust object tracking can be achieved even during extended periods of occlusion when the camera views of trucks are blocked or obscured, a typical phenomenon on highways, especially during rush hours. They were also able to identify the factors that impact detection, which can be used to inform the ideal deployment conditions, e.g., to limit lights directed at the cameras and use appropriate roadway lighting for nighttime sensing.
Because CCTV and WIM sensors do not uniquely identify and track trucks, extracting mobility patterns from their detections is challenging. To estimate truck flows from roadside sensor observations, they propose a novel probabilistic framework that processes the observations to estimate where the truck can be located given past observations and traffic conditions. The team’s experiments on synthetic datasets show that the proposed methods achieve high accuracy while maintaining practical computation time. They applied the proposed framework to several realistically synthesized datasets of roadside sensor observations generated using the truck sensor dataset. The simulations showed that their methods could estimate visit probabilities accurately when sufficient sensors are available, allowing OD matrices to be generated at a fine resolution.
Overall, this research provides a promising direction for sensor-generated truck flow data, addressing a gap in metropolitan transportation planning data. For more information, see the final report from the recent Pacific Southwest Region University Transportation Center research, available to read here.