The research objective of this work was to understand pedestrians’ behavior and interaction with vehicles during pre-crash scenarios that provides critical information on how to improve pedestrian safety.
In this study, we recruited 110 cars and their drivers in the greater Indianapolis area for a one year naturalistic driving study starting in March 2012. The drivers were selected based on their geographic, demographic, and driving route representativeness. We used off-the-shelf vehicle black boxes for data recording, which are installed at the front windshield behind the rear-view mirrors. It records high-resolution forward-view videos (recording driving views outside of front windshield), GPS information, and G-sensor information.
We developed category-based multi-stage pedestrian detection and behavior analysis tools to efficiently process this large scale driving dataset. To ensure the accuracy, we incorporated the human-in-loop process to verify the automatic pedestrian detection results. For each pedestrian event, we generate a 5-second video to further study potential conflicts between pedestrians and vehicle. For each detected potential conflict event, we generate a 15-second video to analyze pedestrian behavior.
We conduct in-depth analysis of pedestrian behavior in regular and near-miss scenarios using the naturalistic data. We observed pedestrian and vehicle interaction videos and studied what scenarios might be more dangerous and could more likely to result in potential conflicts.
We observed: 1) Children alone as pedestrians is an elevated risk; 2) three or more adults may be more likely to result in potential conflicts with vehicles than one or two adults; 3) parking lots, communities, school areas, shopping malls, etc. could have more potential conflicts than regular urban/rural driving environments; 4) when pedestrian is crossing road, there is much higher potential conflict than pedestrian walking along/against traffic; 5) There is an elevated risk for pedestrians walking in road (where vehicles can drive by); 6) when pedestrians are jogging, it is much more likely to have potential conflict than walking or standing.; and 7) it is more likely to have potential conflict at cross walk and junction than other road types.
Furthermore, we estimated the pedestrian appearance points of all potential conflict events and time to collision (TTC). Most potential conflict events have a TTC value ranging from 1 second to 6 seconds, with the range of 2 seconds to 4 seconds being associated with highest percentages of all the cases. The mean value of TTC is 3.84 seconds with standard deviation of 1.74 seconds.
To date, we have collected about 65TB of driving data with about 1.1 million miles. We have processed about 50% of the data. We are continuously working on the data collection and processing. There could be some changes in our observation results after including all data. But the existing analysis is based on a quite large-scale data and would provide a good estimation.