| dc.description.abstract | Understanding the category of a vehicle is an essential study for transportation safety and operation. With the explosive number of GPS devices, there are massive vehicle GPS trajectory data sets whose sizes are beyond the traditional trajectory analysis method's capability. This study utilizes Apache Spark™ to build up a framework whose output data can be compatible with machine learning algorithms for vehicle category classification. Five types of features were extracted from the GPS trajectory data, namely driving habits statistics, trajectory sample quality statistics, geographical information statistics, origin and destination cluster statistics, and temporal statistics. The spatial clustering algorithm and spatial join are incorporated in the workflow, significantly broadening the number of features for the training data set. The results show that the five types of statistics extracted from the trajectory are adequate for distinguishing different vehicle categories by machine learning algorithms. The same accuracy rank sequence for the vehicle classes was observed across different types of features and algorithms, and the decision tree ensemble algorithms have better performance over the logistic regression and support vector machine algorithms. | en |
