TQAM: Temporal Attention for Cycle-wise Queue Length Estimation using High-Resolution Loop Detector Data

2021
Queue Length Estimation along urban arterials is vital to city traffic planners for calculating ‘Level of Service’ measures and optimizing signal plans to alleviate congestion. While several data sources such as GPS, Video, Bluetooth, DSRC etc. are now available, we focus exclusively on high-resolution loop detector data, which is widely available to traffic authorities in North America and elsewhere. Analytical methods for queue length estimation rely on counting input-output vehicle flows over advance and stop-bar detectors or identifying breakpoints in the detector actuation waveforms. However, these methods are limited, as they are sensitive to assumed traffic parameters, detector placement and do not take into account driving behaviors and the effect of left-turn buffers. More recently, Machine Learning models have been developed which learn queue lengths from traffic state data. These are usually trained on localized datasets at coarse time resolutions. In this work, we show that Deep Neural Networks are capable of directly learning an abstract representation of the queuing process, from detector actuation waveforms along an intersection approach with an exclusive left-turn buffer. Using a microscopic traffic simulator, we generate a large dataset by approximately replicating traffic arrival patterns from a realworld loop detector dataset. We then feed multi-cycle data to compute maximum and residual queue lengths across a range of traffic conditions, at the cycle-level. We explore the use of lightweight Encoder-Decoder architectures with Temporal Attention, trained using teacher-forcing strategy, and show their superior performance over regular Feed-Forward Neural Networks.
    • Correction
    • Source
    • Cite
    • Save
    26
    References
    0
    Citations
    NaN
    KQI
    []
    Baidu
    map