ATTACC: Attention Based Accident Anticipation

A Monocular Depth Enhanced Approach

ATTACC is an attention-based accident anticipation system for dashcam-style driving videos. It processes each frame with an object detector and a monocular depth estimator, then fuses the object-centric appearance features with their estimated 3D layout. A graph/transformer core attends over traffic participants and their interactions across time to predict, for every frame, the probability that an accident will occur in the near future. This multi-modal design helps the model focus on high-risk agents and improves early anticipation. The current implementation targets the CCD dataset and is trained with an uncertainty-guided loss (TUGL).

Tech stack and project status

• Language: Python (>= 3.10)
• Package manager: Poetry (pyproject.toml with package-mode disabled)
• Deep learning: PyTorch 2.5.x, torchvision, torchaudio
• GNN: torch-geometric
• Computer vision: OpenCV, matplotlib
• Utilities: numpy, tqdm, scikit-learn
• Optional: CUDA (11.8/12.1/12.4 supported via conda channels in the commands below)

Note: There is no CLI entry point defined in pyproject.toml. Scripts are run as Python modules or by executing files directly. Some paths and hyperparameters are currently hard-coded in scripts.

Requirements

• Python 3.10+
• Conda (recommended) for managing the PyTorch + CUDA stack
• Poetry for dependency management (package-mode=false; deps are pinned in poetry.lock)
• ffmpeg installed on your system (required by matplotlib FFMpegWriter for demo rendering)
• GPU with CUDA (optional but recommended). CPU-only is supported with lower performance.

Environment setup

Follow the steps below to set up the project environment exactly as configured.

1) Create and activate Conda environment

conda create -n attacc python=3.10
conda activate attacc

Note: PyTorch announced that version 2.5 will be the last release published to the pytorch channel on Anaconda.

2) Install PyTorch

macOS (CPU):

conda install pytorch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 -c pytorch

Linux/Windows (choose your CUDA):

CUDA 11.8

conda install pytorch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 pytorch-cuda=11.8 -c pytorch -c nvidia

CUDA 12.1

conda install pytorch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 pytorch-cuda=12.1 -c pytorch -c nvidia

CUDA 12.4

conda install pytorch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 pytorch-cuda=12.4 -c pytorch -c nvidia

CPU only

conda install pytorch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 cpuonly -c pytorch

3) Install Poetry

pip install poetry

4) Install project dependencies via Poetry

This uses the existing pyproject.toml and poetry.lock.

poetry install

5) Install timm separately (to avoid upgrading PyTorch)

pip install timm==0.6.12

Project structure

High-level directories of this repository:

• data/
  • datasets/ccd/
  • preprocessing/
• models/
  • architecture/
  • saved_models/
• outputs/
  • training/
  • testing/
• demo/
• logs/
• visualization/
• pyproject.toml, poetry.lock

Notes and troubleshooting

• torch-geometric may require a PyTorch build that matches your CUDA runtime; if you hit install/runtime errors, verify the PyTorch/CUDA pairing and consult torch-geometric install docs.
• If matplotlib animation raises an ffmpeg error, ensure ffmpeg is installed and on PATH (see setup step 6).
• Some dataset and feature paths are currently hard-coded for CCD; adapt as needed.

TODO

• add argparse to trainer/evaluator/preprocess to make CLI usage consistent and configurable.
• add argparse-based CLIs and [project.scripts] entry points.
• Add citation information and license