In traffic safety studies, there are almost inevitable concerns about unobserved heterogeneity. As a feasible alternative to current methods, this article proposes a novel crash count model that can address asymmetry and multimodality in the data. Specifically, a Bayesian random parameters model with flexible discrete densities for the regression coefficients is developed, employing a Dirichlet process prior. The approach is illustrated on the Ontario Highway 401, which is one of the busiest North American highways. The results indicate that the proposed model better captures the underlying structure of the data compared to conventional models, improving predictive power examined based on pseudo Bayes factors. Interestingly, the model can identify sites (highway segments, intersections, etc.) with similar risk factor profiles, those that manifest similarity in the heterogeneous effects of their site characteristics (e.g., traffic flow) on traffic safety, providing useful insight towards designing effective countermeasures.