UltraMovelets: efficient movelet extraction for multiple aspect trajectory classification

Several methods for trajectory classification build models exploring trajectory global features, such as the average and the standard deviation of speed and acceleration, but for some applications these features may not be the best to determine the class. Other works explore local features, applying trajectory partition and discretization, that lose important movement information that could discriminate the class. In this work we propose a new method, called Movelets, to discover relevant subtrajectories without the need of a predefined criteria for either trajectory partition or discretization. We extend the concept of time series shapelets for trajectories, and to the best of our knowledge, this work is the first to use shapelets in the trajectory domain. We evaluated the proposed approach with several categories of datasets, including hurricanes, vehicles, animals, and transportation means, and show with extensive experiments that our method largely outperformed state of the art works, indicating that Movelets is very promising for trajectory classification.