Network-less trajectory imputation

Abstract

The ability to collect large numbers of trajectory data through GPS-enabled devices have enabled a myriad of very important applications that are widely used on a daily basis. This includes urban computing, transportation, and map APIs for routing and navigation. Unfortunately, a major hinder for all these applications is the accuracy of collected trajectories. Due to low sampling rates, trajectories are usually sparse in terms of the large spatial and temporal distances between each two consecutive collected points. This paper presents TrImpute; a novel framework for trajectory imputation that inserts artificial GPS points between the real ones in a way that the imputed trajectories end up to be very similar to the case if such trajectories were collected with a much higher sampling rate. Unlike all prior trajectory imputation techniques, TrImpute does not assume the knowledge of the underlying road network. This makes it more practical when the underlying road network is not available or inaccurate. Experimental results on real datasets and a real deployment of TrImpute show that it is highly scalable, accurate, and can significantly boost the performance of trajectory applications by feeding them highly accurate trajectories.