Data underlying the MSc Thesis: Toward Occlusion Capable Human Trajectory Prediction

doi:10.4121/3dc88884-d8f4-42db-b643-e799fe7fb432.v1
The doi above is for this specific version of this dataset, which is currently the latest. Newer versions may be published in the future. For a link that will always point to the latest version, please use
doi: 10.4121/3dc88884-d8f4-42db-b643-e799fe7fb432
Datacite citation style:
Féry, Paul (2025): Data underlying the MSc Thesis: Toward Occlusion Capable Human Trajectory Prediction. Version 1. 4TU.ResearchData. dataset. https://doi.org/10.4121/3dc88884-d8f4-42db-b643-e799fe7fb432.v1
Other citation styles (APA, Harvard, MLA, Vancouver, Chicago, IEEE) available at Datacite
Dataset

Here are the dataset and model files related to the MSc thesis: Toward Occlusion Capable Human Trajectory Prediction.

This thesis focuses on handling occlusions and partially missing positional information of agents when predicting their trajectories.


Dataset files comprise three distinct versions of the trajectory dataset used throughout the thesis project:

  • train, val and test splits for the trajectory dataset under fully observed conditions
  • train, val and test splits for the trajectory dataset under occluded conditions
  • test split for the trajectory dataset under occluded conditions, with imputation of missing positions by means of interpolation and constant velocity extrapolation.

Occluded trajectories were generated by applying a simulator of occlusion events onto a publicly available trajectory dataset: the Stanford Drone Dataset. Our dataset Files are therefore derived from the Stanford Drone Dataset.


Model files contain checkpoints with weights that can be used to initialize prediction models from our implementation. These checkpoints are accompanied by some metadata, with information about the evolution of train and validation losses throughout the training process of individual model instances. Models are trained in two separate phases (I, and II): each model file contains all relevant model data for both phases of one model instance.


The source code related to these files is hosted on this GitHub page. The repository's README contains a comprehensive set of instructions on how to use the files in order to reproduce the results we obtained in our research (please refer to the section titled "Downloading Models and Legacy Datasets").

history
  • 2025-01-24 first online, published, posted
publisher
4TU.ResearchData
format
Dataset files: gzipped collections of HDF5 dataset objects. Model files: gzipped directories containing checkpoints saved as python pickle objects. Additional .csv files are present, containing information about train/val losses.
organizations
TU Delft, Faculty of Mechanical Engineering

DATA

files (18)