Traffic conflicts serve as critical proxies for rare crash events in transportation safety research, providing a proactive and data-rich alternative to traditional crash-based evaluations. Through Surrogate Safety Measures (SSMs), we quantify risk and detect near-crash events using continuous traffic flow data. Our approach applies advanced machine learning techniques to identify anomalous conflicts, filtering out noise to reveal safety-critical patterns. To connect conflict detection with crash prediction, Extreme Value Theory (EVT) is utilized to model the tail distributions of conflict indicators. By integrating EVT with conflict samples, we establish a scalable framework to estimate crash frequency across diverse driving contexts and safety scenarios.

Streetview image data offers valuable insights into both the objective characteristics of the built environment and individuals' subjective perceptions of streetscapes. While the influence of the built environment on traffic safety is well-explored, the relationship between human perceptions of streetscapes and traffic crashes remains underexplored. This study employs deep learning methods and a Bayesian multivariate Poisson-lognormal model to inform safety considerations at the micro level in the planning process for the targeted streets. The findings are expected to facilitate safety-focused street planning, and contribute to the development of human-centric cities.

Advanced statistical models were developed to estimate the crash data. Previous research includes analyzing the safety effects of the interaction between commercial vehicle percentage and roadway attributes using tobit regression, temporal instability of truck volume composition on non-truck-involved crash severity using uncorrelated and correlated grouped random parameters binary logit models with space-time variations, establishing safety performance functions for the pedestrian space and the road space using the random-parameter negative binomial regression, and safety effects of real-time weather on accident clearance time, etc.

Crash prediction models are often subject to excessive zero observation because of the rare nature of crashes. To address the problem of imbalanced crash data, a deep generative approach - augmented variational autoencoder, and a Synthetic Minority Over-Sampling Technique for panel data (SMOTE-P) to resolve the excess zero problem or disaggregate analysis of bus involved crashes were proposed to generate synthetic crash data to measure the association between crash and plausible explanatory factors.