Using Machine Learning to Reduce False Alarm in Alerts Generated by Automatic Incident Detection Systems
(Room Palms Ballroom - Booth 100)
16 Jul 19
5:30 PM
-
7:00 PM
Automatic Incident Detection (AID) technologies have been used in the past by Traffic Management Agencies to aid Incident Management. However, surveys have shown that the industry response has been lukewarm, primarily because of the high false alarms rates. The advent of mobile phones and crowdsourced data collection from smartphone applications have significantly increased detection efficiency and reduced time-to-detection, although incident management operations are largely dependent on human operators standing by to respond to phone calls (511) from motorist. On the other hand, Video-based AID has evolved rapidly in the last several years with significant improvements in video quality and computing resources, with a substantially improved potential for automating the detection process, especially under low volume conditions. However, these AID technologies still struggle to separate out vehicles stopped on the road due to recurrent congestion or pulled over on the shoulder from vehicles stopped due to an incident. Hence, the number of false alarms (or non-critical alarms) remains unmanageably high. This study proposes a machine learning framework for developing consolidation strategies and filters that will eliminate the majority of false and non-critical alarms and associate confidence values with the alerts, thereby allowing operators to choose to focus on higher confidence alerts during busy periods. Clustering and evolution patterns of the appearance of multiple alerts, where the basic alerts are generated by the AID based on traffic anomalies such as stopped, slow moving vehicles, or vehicles moving in the wrong direction, are used to train the machine learning algorithm to separate out potential high-impact incidents from congestion or non-critical related stops and slowdowns. For this study, data over a 16-mile bidirectional segment of roadway, during a 3-month period, was obtained from the AID system. Approximately 10,000 alerts generated by the system were manually reviewed, using the screenshots and video clips provided by the AID software as well as being compared with incident logs from incident management operations. The alerts were classified into three sub-categories: incidents, recurrent congestion, and other alerts. Data from two months are used to train the model and the data from the third month is used for validation. The results indicate a significant potential of the framework in consolidating the AID generated alerts to a small number of high-confidence alerts that can be used by operators for real time incident management operations.