Stop-and-go wave super-resolution reconstruction via iterative refinement

TL;DR

The paper tackles the challenge of recovering fine-grained traffic speed fields from coarse radar/loop sensor data to better analyze stop-and-go waves. It introduces a conditional diffusion denoising framework trained on WaveX, a large paired dataset of low- and high-fidelity measurements, and demonstrates iterative refinement outperforms traditional baselines across WD, RMSE, and MAPE while enabling accurate travel-time and wave-speed estimations. Key contributions include the first application of conditional diffusion to traffic measurement refinement, the WaveX dataset release, and comprehensive validation showing improved reconstruction of congested dynamics. This work offers a cost-effective pathway to leverage existing sensor networks for enhanced traffic analysis and management, with open-sourced models and data to spur further research.

Abstract

Stop-and-go waves are a fundamental phenomenon in freeway traffic flow, contributing to inefficiencies, crashes, and emissions. Recent advancements in high-fidelity sensor technologies have improved the ability to capture detailed traffic dynamics, yet such systems remain scarce and costly. In contrast, conventional traffic sensors are widely deployed but suffer from relatively coarse-grain data resolution, potentially impeding accurate analysis of stop-and-go waves. This article explores whether generative AI models can enhance the resolution of conventional traffic sensor to approximate the quality of high-fidelity observations. We present a novel approach using a conditional diffusion denoising model, designed to reconstruct fine-grained traffic speed field from radar-based conventional sensors via iterative refinement. We introduce a new dataset, I24-WaveX, comprising 132 hours of data from both low and high-fidelity sensor systems, totaling over 2 million vehicle miles traveled. Our approach leverages this dataset to formulate the traffic measurement enhancement problem as a spatio-temporal super-resolution task. We demonstrate that our model can effectively reproduce the patterns of stop-and-go waves, achieving high accuracy in capturing these critical traffic dynamics. Our results show promising advancements in traffic data enhancement, offering a cost-effective way to leverage existing low spatio-temporal resolution sensor networks for improved traffic analysis and management. We also open-sourced our trained model and code to facilitate further research and applications.

Figures (13)

• Figure 1: Stop-and-go waves observed from a standard traffic sensor system (left), a high-fidelity sensor system (middle), and a generative model output (right) using the low spatio-temporal resolution data as input. The examples used here are extracted from Interstate 24 near Nashville, Tennessee in the United States, dated June 3rd, 2024 during the morning peak hours. In these examples, the x-axis represents the local time, and the y-axis represents the mile marker of the freeway. The high-fidelity sensors distinctly highlight the stop-and-go wave patterns, providing a much clearer and detailed view of the traffic flow dynamics.
• Figure 2: Demonstration of the forward (left to right) and inverse (right to left) diffusion process in our problem. Noted that the coarse-grained data is not shown here; refer to Figure \ref{['fig:overview']} for the input speed profile of $\mathbf{r}$.
• Figure 3: Time-space trajectory data diagrams for each included day in the I-24 MOTION: WaveX dataset. Each day of data spans 4.2 miles (y-axis) and 4 hours (x-axis) and is produced with $\sim$ 2-foot positional accuracy and at 10Hz. Days with no data due to e.g. system maintenance are filled with dark grey. (pink text) validation holdout days. (blue text) testing holdout (notable event) days. Days with an asterisk were not used in model training and evaluation due to missing data or other anomalous events, but are released with the dataset.
• Figure 4: Locations of I-24 MOTION camera poles (which provide complete roadway coverage between) and I-24 RDS sensor placement.
• Figure 5: The UNet architecture iteratively refines the denoising process by concatenating the output from the previous step $\mathbf{m}_{t}$ with the upsampled low spatio-temporal resolution data $\mathbf{\tilde{r}}$. This process is conditioned on speed, occupancy, and volume of the low spatio-temporal resolution data. The specific UNet architecture employed for our tasks is depicted in the middle of the figure.