UniTraj: A Unified Framework for Scalable Vehicle Trajectory Prediction

Website | Paper

💡UniTraj allows users to train and evaluate trajectory prediction models from real-world datasets like Waymo, nuPlan, nuScenes and Argoverse2 in a unified pipeline.

🔥Powered by Hydra, Pytorch-lightinig, and WandB, the framework is easy to configure, train and log.

📰 News & Updates

Dec. 2024

• 🔥 UniTraj now supports data selection with TAROT! Try to use less data for improved performance.

Nov. 2024

• Adding AV2 evaluation tools.
• Using h5 format data cache for faster loader.

Sep. 2024

• New website is live! Check it out here.

Jul. 2024

• 🚀 Accepted to ECCV 2024!

Mar. 2024

• 🌐 Launched the official UniTraj Website.
• 📝 Published UniTraj in Arxiv.

🛠 Quick Start

0. Create a new conda environment

conda create -n unitraj python=3.9
conda activate unitraj

1. Install ScenarioNet: https://scenarionet.readthedocs.io/en/latest/install.html

2. Install Unitraj:

git clone https://github.com/vita-epfl/UniTraj.git
cd unitraj
pip install -r requirements.txt
python setup.py develop

You can verify the installation of UniTraj via running the training script:

python train.py method=autobot

The model will be trained on several sample data.

Code Structure

There are three main components in UniTraj: dataset, model and config. The structure of the code is as follows:

unitraj
├── configs
│   ├── config.yaml
│   ├── method
│   │   ├── autobot.yaml
│   │   ├── MTR.yaml
│   │   ├── wayformer.yaml
├── datasets
│   ├── base_dataset.py
│   ├── autobot_dataset.py
│   ├── wayformer_dataset.py
│   ├── MTR_dataset.py
├── models
│   ├── autobot
│   ├── mtr
│   ├── wayformer
│   ├── base_model
├── utils

There is a base config, dataset and model class, and each model has its own config, dataset and model class that inherit from the base class.

Pipeline

1. Data Preparation

UniTraj takes data from ScenarioNet as input. Process the data with ScenarioNet in advance.

2. Configuration

UniTraj uses Hydra to manage configuration files.

Universal configuration file is located in unitraj/config/config.yaml. Each model has its own configuration file in unitraj/config/method/, for example, unitraj/config/method/autobot.yaml.

The configuration file is organized in a hierarchical structure, and the configuration of the model is inherited from the universal configuration file.

Configuration Example

Please refer to config.yaml and method/autobot.yaml for more details.

2. Train

python train.py

3. Evaluation

1. In config.yaml, set the ckpt_path to the path of the trained model and val_data_path to the validation data path.
2. (Optional) In config.yaml, set eval_waymo or eval_nuscenes to True if you want to evaluate the model with Waymo or nuScenes official evaluation tool. (Install waymo-open-dataset and nuscenes-devkit first)
3. Runpython evaluation.py

4. Dataset Analysis

python data_analysis.py

Contribute to UniTraj

Implement a new model

1. Create a new config file in unitraj/config/ folder, for example, unitraj/config/new_model.yaml
2. (Optional) Create a new dataset class in unitraj/datasets/ folder, for example, unitraj/datasets/new_dataset.py, and inherit unitraj/dataset/base_dataset.py
3. Create a new model class in unitraj/model/ folder, for example, unitraj/model/lanegcn.py, and inherit from pl.LightningModule

Dataset Structure

Scenario Metadata

• scenario_id: Unique scenario ID representing a traffic scenario.

Object Trajectories

• obj_trajs: Historical trajectories of objects with the following attributes: [0:3] position (x, y, z) [3:6] size (l, w, h) [6:11] type_onehot [11:33] time_onehot [33:35] heading_encoding [35:37] vx,vy [37:39] ax,ay
• obj_trajs_mask: Valid mask for obj_trajs.
• track_index_to_predict: Index indicating which trajectory should be used as the training sample (provided by the official dataset).
• obj_trajs_pos: The first 3 dimensions of obj_trajs representing the x, y, and z coordinates of the objects.
• obj_trajs_last_pos: The x, y, and z coordinates of the last frame in the historical object trajectories.

Centered Objects

• center_objects_world: World coordinates of the centered objects (used as the training sample).
• center_objects_id: ID of the centered objects.
• center_objects_type: Type of centered objects:
  • 1: Vehicle
  • 2: Pedestrian
  • 3: Cyclist

Map Information

• map_center: World coordinates of the map center.
• map_polylines: Polylines representing the map with the following attributes:
  • [0:3]: Position (x, y, z)
  • [3:6]: Direction (x, y, z)
  • [6:9]: Previous point position (x, y, z)
  • [9:29]: Lane type one-hot encoding
• map_polylines_mask: Valid mask for map_polylines.
• map_polylines_center: Center point of each map polyline.

Future State Predictions

• obj_trajs_future_state: Future state of all the objects.
• obj_trajs_future_mask: Valid mask for obj_trajs_future_state.

Ground Truth Data

• center_gt_trajs: Ground truth trajectories of centered objects, including x, y, vx, and vy coordinates.
• center_gt_trajs_mask: Valid mask for center_gt_trajs.
• center_gt_final_valid_idx: Final valid index of the center_gt_trajs.
• center_gt_trajs_src: Ground truth trajectories in world coordinates.

Additional Metadata

• dataset_name: Name of the dataset (e.g., Waymo, AV2, nuScenes).
• kalman_difficulty: Kalman filter difficulty level of the centered object.
• trajectory_type: Type of trajectory (e.g., straight, turn right, turn left).

---

For citation:

@article{feng2024unitraj,
  title={UniTraj: A Unified Framework for Scalable Vehicle Trajectory Prediction},
  author={Feng, Lan and Bahari, Mohammadhossein and Amor, Kaouther Messaoud Ben and Zablocki, {\'E}loi and Cord, Matthieu and Alahi, Alexandre},
  journal={arXiv preprint arXiv:2403.15098},
  year={2024}
}