Goal-Conditioned Social-LSTM

This is my trial to use goal predictor in the Social-LSTM algorithm for complete normalized pedestrian trajectory prediction (20 frames). In this repository, The goal estimator is from this paper:

@inproceedings{he2021where,
  title={Where are you heading? Dynamic Trajectory Prediction with Expert Goal Examples},
  author={He, Zhao and Richard P. Wildes},
  booktitle = {Proceedings of the International Conference on Computer Vision (ICCV)},
  month = {Oct.},
  year={2021}
}

The aim of this repository

In this repository, I would like to look into the effect of end-point information (ground truth end point and predicted end point) on the pedestrian trajectory prediction. The effect of end-point information will be added to my master's thesis.

Introduction

Goal-conditioned deep learning methods, including Goal-conditioned Social-LSTM method, use ground truth end points during training since the ground truth point during training is know. During test, the goal predictor is used firstly since the end points are unknown. And then trajectory predictor is used afterwards.

In this repository, we use the goal-shift method to endow the complete normalized pedestrian trajectory with end-point information. Note that during training, we all use ground truth end points in the goal-shift method, while the predicted end point generated by goal predictor and ground truth end point are used respectively during test, for the sake of looking into if the predicted goal or ground truth goal can attract the trajectory towards the ground truth end points.

Result

The result of Goal-Conditioned Social-LSTM methods using predicted goal and ground truth goal for complete normalized pedestrian trajectory prediction (20 frames) are shown as follows:

Note that the predicted goal is different from ICCV2021.

Note that Goal-Conditioned Social-LSTM A means using ground truth goal during test; Goal-Conditioned Social-LSTM B means using predicted goal during test

ADE: Average Displacement Error (ADE) is the mean square error (MSE) over all estimated points of a trajectory and true points.

Note that the normalized Social-LSTM is added to see if the goal information can improve the ADE.

Dataset	Goal-Conditioned Social-LSTM A	Goal-Conditioned Social-LSTM B	Social-LSTM
ETH	4.4013	3.8677	1.9927
HOTEL	5.3760	5.2482	1.4552
ZARA1	4.1749	3.7574	1.7175
ZARA2	3.8424	4.1727	1.3038
UNIV	3.9242	4.0282	1.9072
Average	4.3438	4.2148	1.6753

FDE: FInal Displacement Error (FDE) is the distance between predicted final destination and true final destination at the end of prediction period (20 frames)

Note that the normalized Social-LSTM is added to see if the goal information can improve the FDE.

Dataset	Goal-Conditioned Social-LSTM A	Goal-Conditioned Social-LSTM B	Social-LSTM
ETH	6.0305	5.2809	3.5463
HOTEL	6.3336	6.1913	2.4048
ZARA1	5.7317	5.1891	2.9853
ZARA2	5.3070	5.3587	2.2892
UNIV	5.0127	5.0815	3.2343
Average	5.6831	5.4203	2.8920