toolkit

OpenTraj Toolkit

The official implementation for the paper:

OpenTraj: Assessing Prediction Complexity in Human Trajectories Datasets
Javad Amirian, Bingqing Zhang, Francisco Valente Castro, Juan José Baldelomar, Jean-Bernard Hayet, Julien Pettré
Published at ACCV2020: [paper], [presentation]

Dataset Analysis

We present indicators in 3 categories:

• Predictability

  • Conditional Entropy conditional_entropy.py
  • Number of Clusters global_multimodality.py
    • The algorithm implemented here reparametrizes all the trajectories in the dataset using a cubic spline. We can take samples of points at specific times (like t = 1.0) for all the trajectories and then calculate the number of clusters and adjust a Gaussian Mixture Model. The following image describes the method for t = 1.0 on the ETH dataset.

      

• Trajlet Regularity

  • Average Speed
  • Speed Range
  • Average Acceleration
  • Maximum Acceleration
  • Path Efficiency
  • Angular Deviation

• Context Complexity

  • Distance to Closest Approach (DCA)
  • Time-to-Collision (TTC)
  • Local Density

Setup

To set up the code and to run the benchmarking, run the following script:

# create a virtualenv and activate it
python3 -m venv env
source env/bin/activate

# install dependencies
cd [OpenTraj]
pip install -r benchmarking/requirements.txt

# run it!
python toolkit/benchmarking . [output_dir]

# exit the environment
deactivate  # Exit virtual environment