Link to Content Area

Institute of Transportation, MOTCInstitute of Transportation, MOTC

  • small size
  • medium size
  • large size
  • print
  • facebook
  • plurk
  • twitter

Summary of IOT Publications

Title Applying Artificial Intelligence Techniques to Identifying Accident-prone Road Sections: Development of Driving Behavior Analysis and Image Recognition Using Out-of-vehicle Camera Videos
Dept Safety Division
Year 2022
Month 6
Price 550

       In the past, the data while driving was seldom collected, and the proximate cause of the accident was the main factor in determining the cause of the accident, but the proximate cause is not necessarily the main cause of the accident, and the prevention of proximate cause alone may not be effective in preventing the accident. The project aims to shift road safety management from a nearsighted accident view to a comprehensive perspective that includes not only accidents but also dangerous driving events and driving behaviors in the analysis. This report presents the efforts conducted in the first year of a four-year project; the aim is to apply image recognition technology to screen abnormal events through off-road image data analysis and identify high-risk driving behaviors that may cause and increase the risk of accidents.
       The project first interviewed the Vehicle Safety Certification Center to identify the regulation trend of ADAS (advanced driver assistance system) equipment. We then conducted an in-depth interview with nine large-vehicle operators and ADAS equipment vendors, and also a questionnaire survey to 25 bus operators to explore the current needs of ADAS, especially in driver and fleet safety management. Based on literature review and road accident analysis, the study defined three common
dangerous events in highway driving, including insufficient vehicle-following distances, lane drifting, and speeding on the curve or ramp. We then deduced event sequences and subsequences for each event category and defined three levels of event danger level, not dangerous, possibly dangerous, and dangerous, based on the interaction between the subject vehicle and the vehicles around, the absolute distance from the leading vehicle, and the driver response at the onset of the event.
Based on lane marking and car-following distances, the study proposed a conceptual framework that divided the subject vehicle and its surrounding space into 12 grids, 20 meters each; evaluation indicators, such as large vehicle presence, or the distance relative to the subject vehicle, were then specified for each grid. The study then developed image recognition algorithms based on Canny edge detection, Cascade Mask R-CNN, and YOLO. Using approximately 20,000 image frames with ground
truth labeled, the analysis showed that the developed image recognition algorithm had an accuracy of at least 90%; yet, the accuracy may deteriorate to 60–80% on ramps, rainy days, low-speed driving, or to grids with three vehicles or more.
       Based on the developed 12-grid framework and image-recognition algorithms, the study examined 200 trips with 2,531 warnings. Among the warnings, only 102 (or 4%) were dangerous events. The extremely high false-positive rate implies the inappropriateness of using the ADAS warning as the only indicator of dangerous driving behaviors. The study then applied
XGBoost for predicting dangerous driving events. The analysis results showed that the prediction accuracy of the image-based analysis was similar to that of the ADAS. The prediction accuracy was further enhanced when both image and ADAS data were included. Based on spatiotemporal analysis, the study identified the concurrence of forward collision warnings and road accidents in segments with a narrow inner shoulder. Dangerous events associated with experienced drivers were less severe, but yet more severe when the events were associated with drivers with multiple shifts within 72 hours.

Post date 2022-06-30
Count Views:134