| Description |
This study is focused on the identification of crash hotspot locations based on historical crash frequency and spatial patterns, and the use of such information to not only estimate future crash frequencies but also predict detailed spatial crash distributions along a freeway corridor. The proposed method is tested on seven years of crash data from 2010 to 2016 along a section of Interstate 15 in the Salt Lake region. A spatial analytic approach based on Network Kernel Density Estimation (NKDE) is applied to estimate crash density with the idea of incorporating the kernel as a density function based on actual network distance rather than Euclidean distance. The NKDE method identifies linear clusters that are referred to as one-dimensional (line) clusters instead of two-dimensional (area) within networks. A linear regression adds a temporal dimension to the analysis by using the NKDE densities from previous years as predictors of the target year, generating an expected distribution of the predicted crash densities. A simple transformation from density to crash frequency exists such that the results can be presented in terms of expected crash intensity. Finally, the hotspot identification process for the future year is validated using PAI (Predictive Accuracy Index) as a measure of performance and percent overlap between predicted and observed hotspot map. Results of the proposed approach showed that most crash hotspots are located around interchanges and the southern part of the study area, and resulted in a total of 88.2% overlap with hotspots developed with actual data from the predicted year. The NKDE output was also compared against the result of another widely used hotspot analysis: the Gi* approach. The output map comparison showed similarities and differences between the performances of these two methods, and the PAI comparison revealed that both hotspot analyses can be potentially useful in the study area. The performance of the linear regression model in terms of R-squared ranged from 0.41 to 0.94, and Root Mean Square Error (RMSE) ranged between 0.75 and 1.27. When linear regression model was applied separately to different segment types, smaller RMSE values were obtained for curved segments. Overall, the results of this study show that the proposed approach can be potentially useful to identify high-crash roadway segments and predict the frequency of crashes and their spatial distribution. The prediction approach can be extended and applied to the output of the Gi* method using the z-score value instead of crash density. |
