/KalmanBaseline

Baseline Kalman models for trajectory predictions. One is constant velocity LSTM

Primary LanguagePython

Baseline Models

This is a code for baseline production using Kalman filter. It is the implementation of the models presented in : https://arxiv.org/abs/1908.11472

To use this code :

  • Set the parameters, dataset path settings.yaml (The Bicycle model may show training unstabilities, contributions are welcome.)
  • Run train_kalman_predict.py starts the trainning.
  • Enter the name of the trained model in the load_name field of settings.yaml (should be in the form <model>_<dataset>_<id>)
  • Run plot_results.py to plot trajectory samples, estimated position and predictions
  • Run save_results.py to save the prediction results computed on the test set
  • Run stats_results.py to print metrics evaluation, plot covariance matching and error histogram (from the saved results)

#Datasets

NGSIM

From NGSIM website:

From googledrive:

Dataset fields:

  • doc/trajectory-data-dictionary.htm

This Dataset is to be pre-processed with the Matlab function preprocess_data.m (that is a slightly modified version of the one from https://github.com/nachiket92/conv-social-pooling)

Argoverse

Get the data from the Argoverse website:

Put their API from Argoverse in a directory at same depth named "Argoverse"

Fusion

The Fusion dataset option relates to a private dataset that is not available.