Early Warning Methods for Traffic High-risk Events Under Low Penetration of Connected and Autonomous Vehicles

Abstract:

We propose an early warning method for high-risk events of traffic operation under low penetration of connected and autonomous vehicles(CAVs）. Specifically， we first define the concept of traffic entropy， and quantifies the micro driving behavior of individual vehicles as a parameter represented by traffic entropy， which is used to characterize the state of macroscopic traffic flow. And then the traffic entropy is used as the input parameter of the Long Short-Term Memory （LSTM） model to establish the early warning model of high-risk events. The HighD Dataset from German highways was utilized for the empirical analyses. In order to compare the application results under CAVs environment， an autonomous-vehicles scenario and a connected-vehicles scenario were set for the high-risk events and non-risk events extracted from the HighD Dataset. and the effectiveness of the warning of high-risk events under different vehicle permeability was compared. Results show that， the false alarm and missed alarm rates of early warning model with traffic entropy parameters are both reduced. Taking the low-penetration CAVs of 10% as an example， the false alarm and missed alarm rates reduced from 6.18 % and 11.47 % to 1.95 % and 3.12 %， respectively. At the same time， the false alarm and missed alarm rates are only 2.28 % and 3.82 % under the prediction environment.