Properties of Virginia Crashes Directly Observed in Traffic Camera Video
There are several sources of crash data compiled in the United States at the state and national level. These databases, several of which only include police reports, are limited in the number of data elements and crashes cannot be reconstructed. There exists a need for a database collected recently with detailed information for crash reconstruction.
Traffic cameras, operated by states, have the potential to directly observe crashes, providing enough information to directly reconstruct the crash. Correlating these crashes with police reports would allow for the integration of occupant, vehicle, and environment information to analyze cases for injury risk modeling, active safety benefits estimation, and other study designs. The objective of this study was to directly observe crashes found in Virginia Traffic camera video.
The source of video in this work was the Virginia Traffic Information system. Video was captured from 1,263 traffic camera video streams starting on December 17, 2019 at 4:00PM and concluding at 11:59PM on December 31, 2020.
Crash data was also captured from the Virginia Traffic Information System. Traffic data was polled every ten seconds for the duration of the study using a program written in Python. When an event with the text “crash” or “accident” in the description was found for the first time within 800 feet of a traffic camera, the entire hour-long segment for that traffic camera was stored as a candidate video segment. Each hour-long segment was then manually reviewed for crashes. A crash was included in the dataset if the type of crash and moment of crash were clearly discernable.
A total of 13,969 candidate video segments were identified from the Virginia Traffic Information System and manually reviewed for crashes. Despite the relatively infrequent placement and narrow field of view of traffic cameras, some cameras were “in the right place at the right time” and a total of 292 crashes were identified within the traffic camera dataset using this method. Only 2.1% of candidate video segments contained crashes.
The most common crash types identified were road departure, rear-end, angle/sideswipe, and intersection. Forward impact crashes, fixed object, and head-on crashes were less common. Road departure and rear-end crashes occurred in similar proportions to national databases, but intersection crashes were underrepresented and severe and rollover cases were overrepresented.
The extraction of crash information from traffic camera video has not been previously identified in the literature, to the best of the authors’ knowledge. This represents a novel analysis method with the potential of producing accurate, detailed, numeric crash information to aid in the development and evaluation of future vehicle safety systems.
Bareiss, M, Gabler, HC, "Vt-CAST" AAAM Short Communication. 2021. Indianapolis, Indiana.