159 — An LSTM Network for Highway Trajectory Prediction

Altché & de La Fortelle (1801.07962)

Read on 26 January 2018
#self-driving-car  #autonomous-car  #autonomy  #prediction  #LSTM  #neural-network 

One of the main challenges of designing self-driving vehicles lives in the interim between when no cars communicate their anticipated plans and when all cars do. If no one knows what everyone else wants to do, we have the current system. If everyone knows what everyone else wants to do, we (ostensibly) have a harmonious coexistance of all road vehicles. But if some agents are predictable and others are not, what we have is a big mess.

This paper works on the challenge of predicting other cars’ (both autonomous and human-driven) positions in the future, based upon their current position and their recent history.

This time-series challenge lends itself particularly well to an LSTM, since more recent timepoints will be more applicable to the future position of the car than older ones.

The model achieves a remarkable score — on average, it mispredicts the latitudinal position of a vehicle ten seconds in the future by only 70 centimeters, improving upon state of the art algorithms.

On the other hand, this model also encounters delays from time to time — sometimes reaching up to ten seconds. So the utility of this system in the short-term is unclear.