Development of Incident Detection Algorithms Compatible with Low Temporal-Resolution Traffic Detectors and Limited Incident Data
Open AccessWith the emerging connected travel environments in the era of intelligent transportation systems, proactive and responsive traffic and safety management tools are desired. Such tools may include an incident detection platform that can distribute real-time traveler information to raise drivers’ awareness of potential and existing traffic incidents and dispatch timely police and emergency service to contain the post-incident losses and impacts.Most of the incident detection algorithms in the offered platforms, however, require traffic detector data with high collection frequency (e.g., 30- or 60-second interval) to achieve desired detection outcomes. In addition, most of the adopted/proposed incident detection algorithms in the past three decades, especially those implementing machine learning (ML), were framed as fully data-driven approaches; dedicated incident data libraries were thus built demanding manual investigations/screenings on incidents. In reality, high-temporal-resolution traffic detector data and comprehensive incident data with significant spatial-temporal coverage are not readily available. In line with the above limitations, the objective of this research is to investigate the potential of the prevailing traffic detector data (i.e., collected/archived with relatively low frequencies by isolated detector stations) in detecting incidents on the corresponding local roadway segments while using collision data as the only available data regarding incidents to frame and solve the incident detection problem. Accordingly, two incident detection algorithms, being either semi-data-driven or non-data-driven, were developed. The first incident detection algorithm adopts a recurrent neural network while changing with alternative yet available data input and use a balancing problem to gain desirable incident detection outcomes when the temporal resolution of traffic detector data and the availability of incident data are limited. Such two-step approach can leverage the learning capability of artificial intelligence (AI) in interpreting traffic flow dynamics but does not rely on specific ML models for direct incident/non-incident classification. The second incident detection algorithm adopts the conventional flow, density and speed measures as input and partitions the flow-density fundamental diagram (FD) into different regimes. Based on the positions of flow-density points in FD, the associated roadway operations are allocated into these regimes and through characterizing the keeping/changing of regimes accompanied by a speed drop, different incident formation scenarios are defined for incident detection purposes. The two proposed algorithms are then compared to the existing algorithms showing their superiorities given the available traffic detector and incident data. The two proposed algorithms are also compared to each other demonstrating their respective advantages with different objectives and priorities on incident detection.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.