Developing Real-Time Crash Prediction Model using Bayesian Network

Abstract

Crashes are regular occurring phenomena andconsequently seek attention as a lot of casualties, both lives and goods, take place as a result of these phenomena. Significant number of studies have been carried out to identify the proper measures to prevent crashes. Earlier studies in this field include determination of black spots and identifying crash patterns to build prediction models with historical data. Traffic flow variables such as ADT, hourly traffic volume or AADT, design speed limit etc. were used to develop models as a part of re-active traffic management. However, later studies have focused on instantaneous traffic flow to develop crash prediction models in real-time as a part of pro-active traffic management. The concept for predicting crash events in real-time is still relatively new and researchers are paying substantial attention in this field of pro-active traffic management. The basic idea is to develop a real-time crash prediction model to identify crash-prone situation so that proper interventions can be provided in order to improve traffic safety on expressways. Many researchers have already put in considerable efforts to develop an effective crash prediction model with significant amount of accuracy and low false alarm. However, there are some major obstacles currently associated with these type of models that must be resolved in order to convert the theoretical arguments into probable actions in real life scenarios. For instance, the real-time crash prediction models use the different interactions among the variables to predict the existing traffic condition. So the models are built specific to a traffic condition within a specific segment of the expressway. This implies the necessity to have a model that can update itself in real-time with newly available traffic data in order to adapt to the traffic condition where it is being employed. Also given the rare nature of crash events, getting matching crash and traffic sensor data is difficult. Hence, an ideal model should be able to be built with relatively low sample size, must have the ability to update itself in real-time and must be capable to adapt to the traffic conditions. With a view to resolving these problems, this study proposes Bayesian Belief Net (BBN) as a platform to build a real-time crash prediction model. BBN as a probabilistic modelling method, allows to solve most of the major hindrances associated with real-time crash prediction model. This study employs data from Shibuya 3 inbound route throughout the months of March, 2014 to August, 2014 provided by Tokyo Metropolitan Expressway Company Limited. The study route is 11.4 kilometer in length and has densely placed loop detectors that suites the requirements of the study to collect instantaneous traffic flow data. As the study is conducted on basic freeway segments only, so data are collected only from 42 specific loop detectors from the basic freeway segments. A total 5 minutes of data prior to crash events are utilized in this study. The traffic flow variables used are 1 minute aggregated avg. occupancy, 1 minute aggregated avg. speed, cumulative 1 minute vehicle count. Separate data bases for crash data and normal traffic data are created to develop 5 BBN models. Each of the BBN models is developed with variables from each of the minutes of total 5 minutes prior to the crash event. HUGIN Expert, a commercially developed software for BBN model building is utilized in this study. For minutes 1, 2 and 4 prior to the crash events, the generated BBN models identified specific influences of the variables on the crash events. The models are evaluated according to their performances. Evaluation criterion for the models are set to be accuracy, overall accuracy and false alarm rates. Model 1 and Model 2 has an overall accuracy of 70% and 58% with crash accuracy of 43% and 70% respectively. The non-crash accuracy of the Model 1 and Model 2 are found to be 83% and 53% with false alarm rates of 63% and 41% respectively. However, Model 4 is found to have fairly met all the criterion to be the best model with an overall accuracy of 60%, crash accuracy of 26%, false alarm rate of 40% and non-crash accuracy of 76%. The result of this study suggest that the newly developed model can successfully fulfill all the aforementioned requirements