Highway Traffic Accident Forecast Based On Real-Time Data And Its Matching Simulation Test

With the development of high technology, such as computer, information,cybernation, and its application into transportation, highway safety research is undergone towards intelligent, digital and systematic direction. It's very important for improving safety, operation, economic and social benefits of road network system to identify potential traffic accident occurrences by using real-time traffic data provided by Intelligent Transportation System and take measures to prevent their occurrence.It's obvious that traditional traffic accident prediction studies can't reflect effects of short-term turbulence of traffic flow on traffic accident occurrences, because they have relied on analyzing historic long-term data to relate environment (such as geometric conditions, weather ), traffic conditions, vehicle characteristics or driver characteristics to traffic accidents or to estimate traffic accident development tendices. In recent years, attempts were made to develop a crash prediction model based on real-time data, which can reflect short-term turbulence of traffic flow. This thesis studies the framework of real-time traffic accident prediction system and the method of traffic accident prediction based on real-time traffic data.Based on the analysis of four factors leading to traffic accident occurrences, the data flow of real-time traffic accident prediction is given. And the framework of real-time traffic accident prediction system is designed which is composed of the detection and surveillance system, hazardous traffic condition identification system and warning information provision system.In the regard of traffic accident prediction based on real-time traffic data, Determination rules of traffic accident precursors and classification standards of dangerous traffic conditions and normal traffic conditions are given. The classification standard of average distances between dangerous traffic conditions and normal traffic conditions based on Euclid distance and the method of determination of calculation time slices which maximizes the sum of squares of differences between precursors of two traffic conditions are developed.50 traffic accidents and their matching traffic data on a dry road are collected from the simulation software TSIS. Three pattern recognition methods– C-means clustering, Bayesian model based on the minimum error rate and k-nearest neighbour method– were applied in this study to classify the traffic patterns into two groups: the patterns leading to traffic accidents (hazardous traffic conditions) and the patterns that do not lead to traffic accidents (normal traffic conditions). Possible reasons for classification results were discussed and recommendations were made for future studies.