[Lane Change] Multi-Modal Trajectory Prediction of Surrounding Vehicles with Maneuver based LSTMs, IEEE IV, 2018

. Summary

1) uses past trajs of vehicles as input

2) outputs possible trajectories w.r.t 6 maneuvers and the corresponding scores

3) Single RNN is used to maneuver decision probability, Encoder-Decoder based RNN is used for traj. prediction

 

1. Input

. (x, y) coordinate of target and its six neighbors

 

2. Network

[Network Structure]

 

. A single RNN is used to produce prob dist over 6 manuevers

. Encoder-Decoder based RNN is used to produce a traj corresponding to a target maneuver

. The architecture is based on the following equation

X : past trajectory

mi : a maneuver

Y : estimated trajectory

 

 

. P(mi|X) corresponds to a single RNN

. P(Y|mi, X) corresponds to Encoder-decoder based RNN

 

3. Results

. NGSIM dataset is used