Certified Human Trajectory Prediction

Here, you can find the main codes of the paper and instructions on how to run them.

---

Table of Contents

• Installation
• Data Preparation
• Model Weights
• Evaluation
  • Original Baselines Evaluation
  • Certified Results Evaluation
  • Visualization
• Citation

---

Installation

1. Clone the repository:

git clone https://github.com/vita-epfl/s-attack.git
cd s-attack/certified

2. Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate

3. Install required packages:

pip install -r requirements.txt

Requirements:

• Python >= 3.8
• PyTorch >= 1.10
• NumPy, SciPy
• See requirements.txt for complete dependencies

---

Data Preparation

This work uses the TrajNet++ dataset, which is shared between the social-attack and certified projects. The repository includes symbolic links from certified/ to social-attack/ for DATA_BLOCK/ and trajnetplusplustools/ to avoid data duplication.

Expected structure after cloning:

s-attack/
├── social-attack/
│   ├── DATA_BLOCK/
│   │   └── trajdata/
│   └── trajnetplusplustools/
└── certified/
    ├── DATA_BLOCK/              # symlink to ../social-attack/DATA_BLOCK
    ├── trajnetplusplustools/    # symlink to ../social-attack/trajnetplusplustools
    ├── baselines/
    └── ...

---

Model Weights

Baseline models trained on TrajNet++ are required for evaluation. Download the weights from here and extract them to baselines/weights/ in the certified directory.

Note: For the rights to use these baseline models, please consult the original authors.

Weights structure:

certified/
└── baselines/
    ├── weights/
    │   ├── Autobot/
    │   │   ├── Autobot_train.pkl.state
    │   ├── DPool/
    │   │   └── d_pool.state
    │   └── EqMotion/
    │       ├── my_checkpoint_v1.pth.tar

---

Evaluation

1. Original Baselines Evaluation

Evaluate original (uncertified) baseline models on the TrajNet++:

python get_original_results.py

This computes ADE, FDE, and collision metrics for all baseline models. Results are saved in:

results/
├── d_pool/uncertified results.txt
├── autobot/uncertified results.txt
└── eq_motion/uncertified results.txt

Note: By default, evaluation runs on 1000 randomly sampled examples for faster testing. To evaluate on the full dataset, use --max_samples -1.

---

2. Certified Results Evaluation

Generate certified predictions with randomized smoothing:

# Run all three baselines (default)
python get_certfied_results.py

Output metrics:

• ADE / FDE: Average/Final Displacement Error
• collision: Collision rate
• cert_collision: Certified collision rate
• FBD / ABD: Final/Average Bound half-Diameter

Results are saved in:

results/
└── {model}/
    ├── mean - {denoiser} denoiser/
    │   └── sigma {σ} - r {r}.txt
    └── median - {denoiser} denoiser/
        └── sigma {σ} - r {r}.txt

Performance Note: This evaluation is computationally intensive as it tests multiple sigma values across all models. By default, the code evaluates on a smaller random subset of data to reduce runtime. To evaluate only a specific model, use --model_types (e.g., --model_types eq_motion).

---

3. Visualization

Generate comparison plots across baseline models:

python get_comparison_plots.py \
    --models d_pool autobot eq_motion \
    --smoothing_types mean median \
    --x_metric fde \
    --y_metric fbd \
    --target_r 0.1

Arguments:

• --models: Models to compare [default: all models]
• --smoothing_types: Smoothing operators [default: mean, median]
• --x_metric / --y_metric: Metrics to plot (see evaluation metrics above)
• --target_r: Certification radius [default: 0.1]

Plots are saved in:

results/plots/

---

Citation

If you find this work useful, please cite our paper:

@InProceedings{bahari2025certified,
    author    = {Bahari, Mohammadhossein and Saadatnejad, Saeed and Askari Farsangi, Amirhossein and Moosavi-Dezfooli, Seyed-Mohsen and Alahi, Alexandre},
    title     = {Certified Human Trajectory Prediction},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2025},
}

---