Examining Network Segmentation for Traffic Safety Analysis With Data-Driven Spectral Analysis

Abstract

Network segmentation is foundational and critical to traffic safety analysis. Existing approaches to conduct segmentation require engineering judgement and are subject to a lack of standard metrics for assessing segmentation performance. This paper presents a novel methodology for data-driven analytics of crash distribution, crash aggregation, and network segmentation. It provides general solutions to determine optimal segment lengths for rigorous safety analysis, and extends the knowledge of crash distribution and aggregation for innovating segment-based safety analysis. The methodology is based on a redesigned spectral analysis of crash density in the spatial frequency domain (SFD) in which frequency components represent the natural patterns how crashes occur along roadways. By proposing the one-dimensional spatial frequency domain analysis (SFDA), this paper reveals the characteristic of power spectral concentration within the low frequency band for crash distribution. Based on this finding, this paper further proposes the power spectral segment length (PSSL) for determining optimal segment lengths and the power spectral percentage (PSP) for assessing the segmentation performance. Based on those new concepts and inferences, the paper proposes the low-pass filtering (LPF) method that outperforms the sliding window (SW) method, and the improved wavelet-based method that identifies high-risk segments properly. Those new techniques are easy to implement and ready for practical application. This research illustrates that interdisciplinary and innovative analytics combined with high-quality data collected by intelligent transportation infrastructure can reshape the fundamental knowledge and conventional paradigms in traffic safety.