Trajectory interpolation: filling the gaps in movement data using long short-term memory
Topics:
Keywords: GPS trajectories, interpolation, long short-term memory, generative adversarial network
Abstract Type: Paper Abstract
Authors:
Zijian Wan, University of California, Santa Barbara
Somayeh Dodge, University of California, Santa Barbara
,
,
,
,
,
,
,
,
Abstract
Movement tracking devices usually sample the location of entities at mostly regular intervals as sequences of timestamped locations, named trajectories. However, many factors (e.g., battery outage, signal loss) may lead to the interruption of tracking data recording, which results in missing data in trajectory datasets. These missing data, termed gaps in this study, need to be dealt with before further analysis. This study proposes a new trajectory interpolation model that leverages a generative adversarial network (GAN) architecture to predict missing trajectory points. In this model, two modules, namely a generator and a discriminator, work against each other during the training process. Long short-term memory (LSTM) layers are used in both the generator and the discriminator of the constructed GAN to process the trajectory data. In the generator, an encoder-decoder structure is leveraged, in which the encoder reads the two observed trajectory segments surrounding a gap, and then the decoder interpolates what is missing in the middle according to the information provided by the encoder. During the training process, the generator aims at generating interpolation results that are close enough to the ground truth, which might fool the discriminator, while the discriminator aims at not getting fooled by the generator. Following the architecture of InfoGAN, we add a latent code in addition to the noise to not only avoid GAN mode collapsing, but it also helps to deal with multi-modality in trajectory interpolation. In this paper, we describe the model and assess the proposed model using a real-world GPS trajectory dataset.
Trajectory interpolation: filling the gaps in movement data using long short-term memory
Category
Paper Abstract