3IA PhD/Postdoc Seminar #20

High-Resolution Traffic Monitoring with Distributed Acoustic Sensing and AI

Traffic management is one of the main challenges that modern cities are facing. Real-time traffic data are a first-order requirement for a range of critical tasks, like traffic jam detection, travel time estimation and the improvement of infrastructures for safer and more environmentally-friendly transportation. Existing techniques for traffic monitoring are mainly based on roadside cameras and road-embedded loop detectors, which have several limitations and drawbacks such as high deployment and maintenance costs, and low spatial resolution. As a low-cost alternative, we are currently developing a real-time traffic monitoring solution based on Distributed Acoustic Sensing (DAS). DAS turns existing fibre-optic (telecommunication) cables into an array of vibration sensors with metric spatial resolution and a range of over one hundred kilometer. Given that telecommunication fibres are often deployed along existing traffic infrastructures, DAS holds great potential for recording vehicles traffic flows with unprecedented spatial resolution. We propose a Machine Learning (ML) model for estimating the speed of vehicles using DAS data. A major component of the proposed model is based on Continuous Piecewise Affine (CPA) transformations, which allows us to extract the speed as a function of space and time. We demonstrate the efficiency of our approach, by comparing our speed estimates to the ones provided by dedicated sensors. The output of the algorithm (i.e., the speed estimates) is visually presented to the user via a dashboard that includes various data analysis tools, accessible directly from a web browser.

A PINN approach for traffic state estimation and model calibration based on loop detector flow data

We analyze the performances of a Physics Informed Neural Network (PINN) strategy applied to traffic state estimation and model parameter identification in realistic situations. The traffic dynamics is modeled by a first order macroscopic traffic flow model involving two physical parameters and an auxiliary one. Besides, observations consist of (averaged) density and flow synthetic data computed at fixed space locations, simulating real loop detector measurements. We show that the proposed approach is able to give a good approximation of the underlying dynamics even with poorer information. Moreover, the precision generally improves as the number of measurement locations increases. [1] M. Raissi, P. Perdikaris and G. E. Karniadakis, Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations, arXiv preprint arXiv:1711.10561 (2017). [2] M. Raissi, P. Perdikaris and G. E. Karniadakis, Physics Informed Deep Learning (Part II): Data-driven Discovery of Nonlinear Partial Differential Equations, arXiv preprint arXiv:1711.10566 (2017). [3] R. Shi, Z. Mo, and X. Di, Physics-Informed Deep Learning for Traffic State Estimation: A Hybrid Paradigm Informed By Second-Order Traffic Models, Proceedings of the AAAI Conference on Artificial Intelligence (2021).

Open discussion on the two contributions