Using Machine Learning to Identify High Impact Incidents in Automatic Incident Detection (AID) System Generated Alarms

Abstract

Video-based Automatic Incident Detection (AID) technologies have been used in the past by Traffic Management Agencies to aid Incident Management. While significant improvements in video quality and computing resources have substantially improved the potential accuracy and efficiency of video-based AID, AID technologies typically still struggle to separate recurrent congestion-related stoppage of vehicles from incident related stoppages. Consequently, the number of false alarms (or non-critical alarms) remains unmanageably high. This study develops a machine learning framework for developing consolidation strategies to minimize false and non-critical alarms and associates confidence values with the alarms, thereby allowing operators to focus on higher confidence alarms during busy periods. The study first investigates the clustering and evolution patterns of the appearance of alarms over time and space. Then it uses this information in the development of a cluster identification algorithm with both spatial and temporal datasets. Finally, the study develops a method for selection of optimal parameters of the machine learning algorithm to separate the alerts for potential high-impact incidents from the alerts related to congestion and other non-critical stops or slowdowns. The results indicate a massive reduction in non-critical alerts without a significant reduction in detection rate or time-to-detect the incidents. While the framework has been used with a video-based AID system, the framework has been developed with interoperability in mind and has the potential to be applicable across all types of AID systems.