A Framework for Real-time Crash Prediction: Statistical Approach Versus

The attempts to predict crashes on freeways through statistical modeling involving capacity driven measures of traffic flow (e.g., AADT) and road geometry have spanned for more than two decades. However, success in crash prediction involving these static data has so far been limited. In recent times, some researchers made efforts to accommodate the weather conditions and seasonal effects to better predict crashes through time series analysis. So far, these models have shown their incapability to accommodate the ever complex human factors, which, to a great extent, are believed to be directly or indirectly responsible for most of the crash cases. Moreover, the prediction outcome of these models is long-term and mainly used to forecast yearly crash, their seasonal variation, identify black-spots or to assess the impact of road improvement initiatives, etc. However, these models overlook the fact that crash can happen even on geometrically correct roads with excellent weather condition due to human factors. One of the latest additions to the crash prediction modeling is the approach to predict crashes based on real-time traffic data which considers the human factors in macroscopic level from the traffic engineering perspective. The assumption in these models is, instantaneous traffic flow data are indirect representation of the human factors. The concept of real-time crash prediction modeling is gaining momentum due to its proactive nature of application and the growing implementation of ITS, which ensures the future availability of real-time traffic data. However, being at its primitive stage and due to scarcity of past real-time data, the present models are yet theoretical and largely prone to unrealistic data requirements and lack of reliability. Moreover, there has not been any well established norm for modeling approach as well. Some researchers preferred using well known statistical methods while some others opted for artificial intelligence based methods. However, the basic concept in modeling crash in real-time has remained the same, in which, the traffic flow data are separated into two groups – condition leading to crash and normal condition, based on specified assumptions, and then future traffic flow data are evaluated to estimate how likely they are to fit into each of these groups. As crash is considered as a rare event, these models need to have the capability of being frequently updated as soon as new data are available. In future, we can expect that the models will also incorporate nontraffic flow variables to attain higher prediction accuracy. Moreover, the operation of these models is highly dependent on the reliability of the detectors, and hence, is expected to predict even when some data are missing. Considering these requirements, in this paper, we have investigated the applicability of Bayesian Network (BN), a popular real-time prediction method in the field of information science, as a probable method to predict crash in real-time. Bayesian Network is a graphical probabilistic modeling approach which can be calibrated in real-time with limited effort, can handle integration of new variables in future with little effort and also suitable in predicting with missing data. Although its inherent features are suitable for real-time crash prediction, it is important to make sure that BN exhibits satisfactory level of accuracy in predicting crashes, too. For this, after developing the model with BN we developed another model with Binary Logistic Regression and used it as a baseline to investigate the prediction capability of BN.