Skew-Normal Distributions for Modeling Asymmetric Moving Tendencies in Pedestrian Trajectories (Under Review)

This is an exemplar repository developed based on GraphTERN to show the implementation of Skew-Normal Module for Trajectory Prediction. The code can be adapted to other methods that model the output distribution as normal distributions or Gaussian mixture distributions.

File structure

• SN-GraphTERN: code that adapts GraphTERN to SN-GraphTERN by replacing the Gaussian mixture distribution with skew-normal mixture distribution.
• skew_normal_class.py: define key classes and functions on skew-normal distributions.
• my_mixture_same_family.py: modify the MixtureSameFamily class from torch.distributions for more numerically stable implementation of the log-likelihood of mixture distributions.
• generaly.py: provide instrumental functions.
• README.md: provide a brief description of the repository.

Environment

• Ubuntu 24.04
• Cuda 11.7
• Python 3.8
• PyTorch 1.13.1

Scripts for running the experiments

cd SN-GraphTERN
# train
# available datasets: eth, hotel, univ, zara1, zara2, sdd
CUDA_VISIBLE_DEVICES=0 python train.py --dataset dataset --dist skew
# test
CUDA_VISIBLE_DEVICES=0 python test.py --dataset dataset --dist skew --date "YYYY-mm-dd_HH-MM-SS"