296 — A Scalable Framework for Trajectory Prediction

The previous papers I’ve recently read on trajectory prediction (#209, #268) have focused on neural network based machine learning.

This paper instead aims to predict the motion of entities — namely, cars across a map network — using a Markov model based approach, which clusters similar trajectories and then classifies newly observed trajectories into one of these clusters in order to easily begin to approximate its unknown trajectory.

This system, named Traj-clusiVAT, guesses competitively with existing models when run on over 3 million passenger trajectories from a Singaporean taxi GPS dataset — potentially one of the largest trajectory prediction tasks ever.

This implementation is very interesting because it embeds trajectories in a network map rather than in continuous space, which means that the clusters can be represented by a single ‘summary’ path across the network, if I understand this work correctly.

This sort of prediction is key for identifying possible navigation situations before they happen: For example, this could enable traffic prediction before the traffic even became a problem. Or it could improve navigation in systems where there is not a consistent GPS signal avaialble, such as dead-reckoning systems used in “offline mode” GPS.