Real-time crash risk estimation on freeways is the key to proactive traffic management. The estimation relies on binary classification based analysis of loop detector data observed prior to historical crash and non-crash cases. How the non-crash cases are sampled with respect to the crashes has interesting consequences for the application of these crash risk estimates. This study looks at the two procedures for sampling the non-crash cases with respect to the rear-end crashes: randomly selected non-crash data and with-in stratum matched non-crash data. The models developed based on these two strategies are provided along with their strengths and shortcomings. The implications of the sampling schemes when testing various proactive traffic management strategies in a microscopic traffic simulation environment are also discussed. It was inferred that matched sampling based models provide valuable insights into association of traffic speed differential with real-time crashes. However, crash risk estimates from random sampling based models are better suited for a universal framework of real-time crash risk estimation as well as for evaluating various traffic management strategies such as ramp metering and variable speed limits in a micro-simulation environment. (A). For the covering abstract of the conference see E216632.
Samenvatting