Traffic crash prediction based on incremental learning algorithm

Real-time crash prediction models are playing a key role in transportation information system. Support vector machine (SVM), a classification learning algorithm, was introduced to evaluate real-time crash risk. The size of traffic dataset is always large with a high accumulating speed. By applying a warm start strategy, an incremental learning algorithm is introduced to update the original model. In this way, incremental dataset will improve the original model with a little time consumption. This study developed crash risk prediction model utilizing loop detector traffic data and historical crash data. With three comparison experiments, the improvement of accuracy and efficiency of this incremental learning algorithm was proved.