The paradigm of connected vehicles, the progressive integration of smartphones and car infrastructure, the rise of Web 2.0, and the progressive emergence of automation onboard vehicles have created a very fertile ground for GPS data sources from probe vehicles to be collected at an unprecedented scale. Yet, the lack of ubiquitous connectivity, the parsimonious use of bandwidth and smartphone battery limits, and the lack of system reliability often makes this data sparse. In particular, it is very common today for GPS traces to be sampled at very low frequencies (on the order of minutes), leading to challenges in using this data. While some studies predict an 80% penetration of the cellphone market by GPS enabled devices by 2015, it is not clear that this figure will translate into ubiquitous GPS data for traffic information systems. As reported in literature, furthermore, millions of fleet vehicles today produce GPS traces sampled at low frequencies, and the paradigm of connected (and automated) vehicles does not automatically translate into high fidelity GPS traces. We have presented a novel class of algorithms to track moving vehicles on a road network: the path inference filter. This algorithm first projects the raw points onto candidate projections on the road network and then builds candidate trajectories to link these candidate points. An observation model and a driver model are then combined in a Conditional Random Field to infer the most probable trajectories. The algorithm exhibits robustness to noise as well as to the specificities of driving in urban road networks. It is competitive over a wide range of sampling rates (1 second to 2 minutes) and greatly outperforms intuitive deterministic algorithms. Furthermore, given a set of ground truth data, the filter can be automatically tuned using a fast supervised learning procedure. Alternatively, using enough GPS data with no ground truth, it can be trained using unsupervised learning. Experimental results show that the unsupervised learning procedure compares favorably against learning from ground truth data. This algorithm supports a range of trade-offs between accuracy, timeliness and computing needs. It extends the current state of the art in its most accurate setting. The results are supported by the theoretical foundations of Conditional Random Fields. Because no standardized benchmark exists, the authors have released an open-source implementation of the filter to foster comparison with other methodologies using other datasets presented in literature. The authors have written an industrial-strength version in the Scala programming language, deployed in the Mobile Millennium system. This multicore version maps GPS points at a rate of several thousands of GPS points per second for the San Francisco Bay area and other large urban areas.


    Zugriff

    Zugriff über TIB

    Verfügbarkeit in meiner Bibliothek prüfen

    Bestellung bei Subito €


    Exportieren, teilen und zitieren



    Titel :

    The path inference filter: Model-based low-latency map matching of probe vehicle data


    Beteiligte:


    Erscheinungsdatum :

    2013


    Format / Umfang :

    17 Seiten, 10 Bilder, 20 Quellen





    Medientyp :

    Aufsatz (Konferenz)


    Format :

    Print


    Sprache :

    Englisch





    The Path Inference Filter: Model-Based Low-Latency Map Matching of Probe Vehicle Data

    Hunter, Timothy / Abbeel, Pieter / Bayen, Alexandre | IEEE | 2014


    Path Inference Filter and Route Choice Model Aided Map-Matching for Low-Frequency GPS Data

    Xiao, Pinghui / Fang, Jie / Xu, Mengyun et al. | TIBKAT | 2021


    Methodology for Calculating Latency of GPS Probe Data

    Wang, Zhongxiang / Hamedi, Masoud / Young, Stanley | Transportation Research Record | 2017