Classification of Precrash Traffic Characteristics to Identify Freeway Crash Conditions: Is a Pattern Recognition Approach Applicable?

There is a growing interest in the area of linking real-time traffic flow characteristics to the probability of freeway crash occurrence. This study is to explore and identify whether there exist such abnormal pattern changes in traffic flow prior to freeway crash events. A crash data set in a period of 12 months as well as corresponding pre-crash traffic data and matched non-crash traffic data on three interstate highways in Northern Virginia area were collected. Nonparametric statistical tests were conducted to preliminarily identify the significant differences between the sample distributions of traffic flow characteristics under pre-crash and non-crash conditions. Then, a statistical pattern recognition method based on minimum Bayes risk decision rule was performed based on estimated probability density functions of each single variable. Finally, back-propagation neural network (BPNN) was used as a multivariate nonlinear classification procedure to the problem. Different combinations of time windows and prior times were tried to get the best classification performance. However, results from the statistical tests and the two different pattern classification methods based on the data used did not support the hypothesis that there are definite abnormal pattern changes in pre-crash traffic flow characteristics.