GTraclus: A Local Trajectory Clustering Algorithm for GPUs

Due to the high availability of location-based sensors like GPS, it has been possible to collect large amounts of spatio-temporal data in the form of trajectories, each of which is a sequence of spatial locations that a moving object occupies in space as time progresses. Many applications, such as intelligent transportation systems and urban planning, can benefit from clustering the trajectories of cars in each locality of a city in order to learn about traffic behavior in each neighborhood. However, the immense and ever-increasing volume of trajectory data and the concept drift present in city traffic constitute scalability challenges that have not been addressed. In order to fill this gap, we propose the first GPU algorithm for local trajectory clustering, called GTraclus. We present a parallelized trajectory partitioning algorithm which simplifies trajectories into line segments using the Minimum Description Length (MDL) principle. We evaluated our proposed algorithm using two large real-life trajectory datasets and compared it against a multicore CPU version, which we call MC-Traclus, of the popular trajectory clustering algorithm, Traclus; our experiments showed that GTraclus had on average up to 24X faster execution time when compared against MC-Traclus.

Automated summary

With the availability of location-based sensors like GPS, large amounts of spatio-temporal data in the form of trajectories have been collected, leading to opportunities for applications like intelligent transportation systems and urban planning. Researchers proposed the GPU algorithm GTraclus for local trajectory clustering to address scalability challenges in city traffic, achieving up to 24X faster execution time compared to the traditional CPU algorithm MC-Traclus.