Abstract
Vehicle headway, defined as the time elapsed between two successive vehicles passing a roadway point, is a key mesoscopic-scale measure in traffic flow theory with safety-critical transportation applications, such as preemptive collision avoidance warning systems as well as connected and autonomous vehicle (CAV) platoon control. Hence, it is crucial to accurately predict vehicle headway over sufficiently long future horizons (i.e., multi-step-ahead prediction) to be applicable for downstream safety-critical applications. This is a challenging task due to several random factors influencing headway, including inter- and intra-driver heterogeneity, asymmetric car-following driving behavior, and vehicle heterogeneity under mixed traffic of different vehicle classes. This becomes even more complicated under traffic congestion, which results in tangible inter-vehicle interactions and, thus, speed-dependent headways. The complex effects of the above factors on headway, along with the unprecedented amount of high time-resolution vehicle trajectory big data (e.g., datapoints recorded every 0.1 second), call for advanced data-driven headway prediction models. Deep learning architectures, particularly variants of Recurrent Neural Network (RNN), are promising candidates as they can “learn” highly nonlinear relationships from headway time-series data. However, recurrent networks are notorious for the vanishing gradient problem, which precludes learning long-term dependencies in time series data. To tackle, this proposed project will employ a state-of-the-art interpretable deep learning model for multi-step-ahead time series forecasting (e.g., next 5 seconds), which can accommodate reasonably long prediction horizons that can capture human/vehicle reaction time. Leveraging the vehicle trajectory big data from the USDOT’s Next Generation Simulation (NGSIM) dataset, the model will be trained and tested to investigate the effects on headway of microscopic traffic measures, macroscopic traffic flow, vehicle class, and lane position.
Description
Timeline
Strategic Description / RD&T
Page 17 of the USDOT’s RD&T Plan: Table 3: “Data-Driven System Safety”
Deployment Plan
Research with potential deployment by auto manufacturers and traffic agencies.
Expected Outcomes/Impacts
The expected outcome of this research project is the capability enhancement in traffic safety (e.g., preemptive collision avoidance systems) and safe operations of connected vehicles (e.g., vehicle platoon control) through the application of the proposed method in the context of inter-vehicle headway prediction.
The anticipated impacts of this research project are 1) improved computational technology in enhancing vehicle safety in car following scenarios when drivers are not completely attentive to the road environment, which can be of particular interest for auto manufacturers to incorporate it into preemptive collision warning systems embedded in smart vehicles; and 2) enlargement of the pool of trained transportation professionals at the nexus of transportation safety and machine learning.
Expected Outputs
The anticipated output of this research project includes a state-of-the-art machine learning method implemented on the vehicle trajectory big data.
TRID
Over the past decades, several statistical models were proposed to analytically derive closed-form distributions for vehicle headway (e.g., see a 2017 survey in the TRID search). However, accurate analytical modeling of headway distribution is challenging due to a multitude of (un)known factors shaping headway distribution, encompassing 1) inter/intra-driver heterogeneity, indicating that different drivers show diverse perceptual reactions and even the same driver may behave differently over time and space, leading to distinct headway distributions; 2) asymmetric driving behavior, referring to the preference of drivers in maintaining shorter headways, resulting in rightly-skewed headway distributions; 3) vehicle heterogeneity under mixed traffic, meaning that different vehicle classes have distinct performances (e.g., vehicle speed, acceleration), giving rise to irregular headway distributions; and 4) congested traffic conditions, under which vehicle speeds are lower than free-flow speed and inter-vehicle interactions are not negligible, causing the headway distribution to be speed-dependent. The complex effects of the above factors on headway warrant the development of more advanced models that can effectively accommodate them. Deep learning can serve as a promising candidate as it can “learn” highly nonlinear relationships governed by headway “data” rather than imposing a priori known closed-form specification for headway and fitting data to that specification.
Individuals Involved
| Email |
Name |
Affiliation |
Role |
Position |
| fatemeh.nazari@utrgv.edu |
Nazari, Fatemeh |
University of Texas - Rio Grande Valley |
Co-PI |
Faculty - Untenured, Tenure Track |
| h.noruzoliaee@utrgv.edu |
Noruzoliaee, Mohamadhossein |
University of Texas - Rio Grande Valley |
PI |
Faculty - Untenured, Tenure Track |
Budget
Amount of UTC Funds Awarded
$41280.00
Total Project Budget (from all funding sources)
$93280.12
Documents
Match Sources
No match sources!
Partners
| Name |
Type |
| City of Edinburg |
Deployment & Equity Partner Deployment & Equity Partner |
