交通事故頻次、嚴重度及碰撞型態之整合模化

Abstract

事故分析與預測之研究，對於事故肇因之確認及預防，具有相當重要之意義，也吸引相當多的學者投入研究。而且，大多數研究多著重於事故頻次或嚴重度之單獨模化與探討。惟路段（或路口）危險程度係同時由事故發生頻次、嚴重度及碰撞型態所代表，在分析事故肇因及研提改善策略，此三項指標缺一不可。況且，此三者間也存有相互關聯性，為能獲得更深入且正確的分析結果，頻次、嚴重度及碰撞型態實不宜分開模化分析。基此，本研究擬以三個研究年度進行下列研究： 1.第一個研究年期：建立事故時空最佳分段（Optimal segmentation）方式 為避免在進行時空分割時導致之時空自我相關問題(Spatiotemporal autocorrelation)，本研究擬利用遺傳演算法（Genetic algorithms, GA）進行高速公路事故資料之最佳化切割，其目標函數可設定為兩個：(1)被解釋變數（事件頻次）之自我相關性最小化，以及(2)同一分割樣本之同質性最大化。本研究將分別將依不同目標函數所分割得之樣本資料，重新進行聯合模式之推估，並與前述推估參數進行比較與檢定，藉此歸納有限混合模式的考量變數。 2.第二個研究年期：建立事故頻次、嚴重性及碰撞型態之聯合推估（Joint estimation）模式 分別考慮事故嚴重程度模式（次序普羅比、多項羅吉特及混合羅吉特）及事故頻次模式（卜瓦松、負二項迴歸、卜瓦松-對數常態）之聯合推估模式。例如，結合多項羅吉特及負二項迴歸之模式，即可稱為多項卜瓦松模式（Multinominal Possion model, MPM）。鑑於兩模式之殘差互有關聯，所以必須就其參數推估之理論公式重新加以推導。就推導所得之公式，本研究擬利用最大概似法（Maximum likelihood estimation, MLE）或最大模擬概似法（Maximum Simulated Likelihood Estimator, MSLE）與研究其它可能更具效率的演算法(Frühwirth-Schnatter, 2006)，如EM（Expectation-Maximization）或Bayesian等方法進行有關參數推估。 3.第三個研究年期：建立事故頻次、嚴重性及碰撞型態之有限混合模式（Finite mixture model, FMM） 本研究將進一步依據事故實際發生之後果，包括死亡人數、受傷人數及財損額度，搭配事故類型（單車、兩車及多車），進行有限混合模式之理論推導與參數推估，據以實證最據解釋能力之事故嚴重度分類方式（Latent class segmentation），並對於我國的肇事分類方式進行檢討。

Numerous studies have been devoted to exploring the factors affecting crash frequency and severity, in order to improve highway safety. However, most of literatures identify the risk factors of crash accounting for either frequency, severity or collision type, separately. To identify high-risk roadway segments along with their contributory factors, obviously, crash frequency, severity, and collision types are three key indices to represent the risk of a roadway segment and to propose safety improvement strategies. Greater scope and in-depth insights can not be obtained without simultaneous consideration of them. Based on this, this study aims to propose an optimal spatiotemporal segmentation model based on genetic algorithms to minimize the spatial autocorrelation (in terms of Moran’s I index) and series autocorrelation. With the segmented samples, this study attempts to develop a joint estimation model of crash frequency, severity and collision type in the first year. At last, based on the research results and experiences of the first two years, this study attempts to develop a more flexible model of crash frequency, severity, and collision type, a finite mixture model, to consider the potential class variables, such as crash severity and collision types for a better investigation on the distribution of crash frequency. With application to freeway and city street accident data, the applicability and performance of the proposed models are demonstrated with corresponding safety improvement strategies proposed.