DAS and machine learning

Vehicle Weight and Speed Recognition with DAS

Connecting roadside observations with fibre-optic waveforms

Distributed acoustic sensing turns an optical fibre into a dense array of vibration measurements. Vehicle passages produce coherent trajectories in time–distance records, but using those signals for vehicle characterization requires reliable alignment between the DAS wavefield and independently observed vehicles.

Distributed acoustic sensing record of several vehicle passages — A data-association and waveform-analysis workflow for estimating vehicle motion and investigating weight-sensitive signatures in distributed acoustic sensing records.

Roadside image of trucks used as visual vehicle labels — A data-association and waveform-analysis workflow for estimating vehicle motion and investigating weight-sensitive signatures in distributed acoustic sensing records.

Data: DAS traces and roadside imagery
Targets: Vehicle speed and weight
Contribution: Matching, labelling, and waveform analysis

Building labelled observations

I developed a workflow to match manually captured vehicle images with the corresponding DAS traces. The association links a visible vehicle passage to its trajectory across sensing channels, providing labels that are otherwise absent from the fibre-optic data.

Accurate matching is essential because traffic can be dense and several trajectories may overlap. Timing, travel direction, lane position, and the slope of the DAS trajectory together constrain which waveform belongs to each observed vehicle.

Roadside camera frame showing trucks travelling along the monitored road — Roadside imagery provides an independent visual record for associating individual vehicles with contemporaneous DAS signatures and constructing labelled examples.

Reading motion from the DAS wavefield

In a time–distance representation, a moving vehicle generates an inclined coherent arrival. Its slope provides a direct constraint on speed, while the spatial continuity of the trajectory helps separate a vehicle signal from stationary or incoherent noise.

The waveform amplitude and spectral content may also contain information related to axle load and total vehicle weight. Those observables are influenced by additional factors—including vehicle type, lane distance, pavement response, and fibre coupling—so weight estimation requires calibration and uncertainty analysis rather than a single amplitude threshold.

Time-distance DAS record with sloping vehicle trajectories — DAS time–distance record containing several vehicle passages. Coherent sloping arrivals trace motion along the road; their slopes constrain speed, while waveform and spectral attributes provide candidate features for weight estimation.

Current contribution and scope

My contribution focuses on the data pipeline needed before model training: synchronizing visual observations with DAS data(though limited data due to random time shift of camera), extracting trajectories, assigning labels, and examining waveform features that may carry weight information.

The present work supports the development of a more reliable estimation model, but the supplied material does not yet establish a final accuracy or a deployment-ready system. Robust evaluation will require held-out traffic observations and explicit tests across vehicle types, lanes, environmental conditions, and fibre-coupling states.