Crowd modeling with Denoising Diffusion Probabilisitic

This project explores the use of Denoising Diffusion Probabilistic Models (DDPMs) to generate sequences of crowd macroproperties—aggregated features representing the movement and behavior of pedestrian groups over time.

---

📂 Dataset

The full ATC dataset contains 41 files, each around 1GB in size.
To facilitate training and quick experimentation, a subset of the dataset is provided in datasets/ATC/pickle/

This subset includes 30-minute pedestrian sequences described by its pre-processed macroscopic properties, such as density, velocity, and flow.

---

⚙️ Installation

Before training or sampling, install the required dependencies using the provided requirements.txt:

> pip3 install -r requirements.txt

Training

To train the model using the provided data subset, simply run:

> python3 traing.py

The training configurations are already set up to work with the subset provided.

Sampling example

Once a model is trained, you can generate new samples from it using the script generate_samples.py. This script allows you to visualize the generated macroproperty sequences through both static and dynamic plots.

Follow the next steps to generate macroproperties sequence samples:

1. Download the Trained Weights
   Download the pretrained model weights from the link below, trained using the full dataset: Download model weights.

   Place the downloaded weights inside the saved_models/ directory at the root of the repository.

2. Generate Samples
   You can choose between static or dynamic visualizations, these visualizations help assess how well the diffusion model captures and predicts the dynamics of crowd macroproperties over time:

   🖼️ Static Plot Example

   Run the following command to generate a static visualization of predicted macroproperty sequences. It compares the prediction with the ground truth (GT), shown row by row.

   python3 generate_samples.py --plot-type="Static" --plot-mprop="Density&Vel" --plot-past="Last2" --vel-scale=0.25

   

   🖼️ Dynamic Plot Example

   Run this command to generate a dynamic animation of the predicted macroproperties. The animation displays past frames followed by predicted frames, marked with black and blue labels for clarity at the bottom.

   > python3 generate_samples.py --plot-type="Dynamic" --vel-scale=0.25 --vel-unc-scale=3

   Below is an example from one of the generated sequences: