Connected Vehicle Trajectory–Based Estimation of Cycle-by-Cycle Traffic Volume: A Deep Learning–Based Approach Considering Slowed Vehicles and Spatiotemporal Shockwave Dissipation

Many existing approaches for estimating cycle-by-cycle traffic volume from vehicle trajectories require either a high market penetration rate of sampled trajectories or auxiliary intersection data (e.g., detector counts or calibrated parameters), largely because they rely only on queued vehicles. In contrast, the motion of nonqueued vehicles—particularly slowed vehicles approaching a dissipating queue—contains informative spatiotemporal cues about queue dynamics. This paper proposes a deep learning–based approach that uses slowed connected vehicles (CVs) to estimate cycle-by-cycle traffic volume under low market penetration. Using a prior-trajectory data set with signal timing information only, a multilayer perceptron (MLP) learns (1) a spatiotemporal capacity-state speed surface, (2) the speed of the last queued vehicle, and (3) the queued-vehicle position (queue length). In operation, the trained models use one slowed probe vehicle per cycle to infer queue length and compute cycle volume without detector counts or externally calibrated shockwave parameters. In simulation (30 representative cycles; 10% market penetration), the method achieved 88.2% accuracy [mean absolute percentage error (MAPE) = 11.8%]. On real-world Federal Highway Administration (FHWA) Next Generation Simulation (NGSIM) trajectories, the method achieved 76.43% accuracy (MAPE = 23.6%). Overall, the results indicate that acceptable volume estimates can be obtained at market penetration rates at which existing approaches typically require higher penetration or additional data sources.