MIAT: Maneuver-Intention-Aware Transformer for Spatio-Temporal Trajectory Prediction

This repository contains the codebase of MIAT.

Paper: MIAT: Maneuver-Intention-Aware Transformer for Spatio-Temporal Trajectory Prediction
Authors: Chandra Raskoti, Iftekharul Islam, Xuan Wang, Weizi Li
Institution: University of Tennessee, Knoxville; George Mason University

Abstract

Accurate vehicle trajectory prediction is critical for safe and efficient autonomous driving, especially in mixed traffic environments when both human-driven and autonomous vehicles co-exist. However, uncertainties introduced by inherent driving behaviors—such as acceleration, deceleration, and left and right maneuvers—pose significant challenges for reliable trajectory prediction. We introduce a Maneuver-IntentionAware Transformer (MIAT) architecture, which integrates a maneuver intention awareness control mechanism with spatiotemporal interaction modeling to enhance long-horizon trajectory predictions. We systematically investigate the impact of varying awareness of maneuver intention on both shortand long-horizon trajectory predictions. Evaluated on the real-world NGSIM dataset and benchmarked against various transformer- and LSTM-based methods, our approach achieves an improvement of up to 4.7% in short-horizon predictions and a 1.6% in long-horizon predictions compared to other intentionaware benchmark methods. Moreover, by leveraging intention awareness control mechanism, MIAT realizes an 11.1% performance boost in long-horizon predictions, with a modest drop in short-horizon performance. The source code and datasets are available at https://github.com/cpraskoti/MIAT.

Key features:

• Unified Architecture: Integrates maneuver intention awareness directly into the Transformer learning objective.
• Tunable Control Mechanism: Allows explicit control over the trade-off between minimizing trajectory prediction error and correctly classifying the underlying maneuver.
• Performance: Achieves significant improvements in long-horizon predictions (up to 11.1%) on the real-world NGSIM dataset compared to intention-aware benchmarks.

Project Structure

MIAT/
├── config.yaml          # Configuration file for experiments
├── requirements.txt     # Python dependencies
├── src/                 # Source code
│   ├── config.py        # Configuration loader
│   ├── data_loader.py   # Dataset classes (NGSIM, HighD)
│   ├── model.py         # MIAT model architecture (Encoder, Decoder, Generator)
│   ├── train.py         # Training script
│   └── evaluate.py      # Evaluation script
├── data/                # Data directory
│   ├── processed_data/  # Processed .mat files for training
│   └── NGSIM_raw_data/  # Raw NGSIM data
└── docs/                # Documentation and paper

Installation

1. Clone the repository.

git clone https://github.com/cpraskoti/MIAT.git

2. Create a virtual environment

python -m venv venv

Activate environemt for Unix

  source venv/bin/activate

Activate environemt: for windows

  venv\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

Data Preparation

Download processed dataset from GOOGLE DRIVE OR BAIDU (4p44)

Save these processed .mat files containing in /data/processed_data/ folder.

• Train Set: data/processed_data/TrainSet.mat
• Validation Set: data/processed_data/ValSet.mat
• Test Set: data/processed_data/TestSet.mat

See src/Data_description.md for detailed information on the data dataset. See src/data_processing.md for detailed information on how data was processed.

Configuration

The project uses config.yaml for experimental settings. You can modify this file or override parameters via command line arguments.

Key Parameters:

• experiment.name: Name of the experiment (used for logging).
• model.encoder_size: Dimension of the transformer model.
• model.use_maneuvers: Whether to use maneuver classification.
• training.batch_size: Batch size.
• data.dataset: "ngsim"

Usage

Training

To train the model with default configuration:

python3 src/train.py

To override configuration parameters:

python3 src/train.py --exp_name my_experiment --gpu_id 0

Training logs and checkpoints will be saved in checkpoint/.

Evaluation

To evaluate the best trained model:

python3 src/evaluate.py

This will compute RMSE or NLL metrics on the test set.

Extensions & Future Work

This codebase provides a solid foundation for further research in trajectory prediction. Here are some areas for potential improvement and extension:

1. Multi-modal Distribution: Extend the model to generate multi-modal trajectory distributions explicitly (e.g., using Gaussian Mixture Models or CVAE) to capture a wider range of plausible future paths.
2. Dataset Scaling: Validate the robustness of MIAT on larger, more complex datasets like Argoverse, NuScenes, or Waymo Open Motion Dataset.
3. Graph Neural Networks: Replace the spatial interaction mechanism with more advanced Graph Neural Networks (GNNs) or Graph Attention Networks (GATs) to better model agent-to-agent interactions.
4. Ablation Studies: Experiment with different weighting factors for the intention awareness control mechanism in config.yaml (scale_cross_entropy_loss) to further analyze the trade-off between maneuver classification and trajectory accuracy.

Citation

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

@article{raskoti2025miat,
  title={MIAT: Maneuver-Intention-Aware Transformer for Spatio-Temporal Trajectory Prediction},
  author={Raskoti, Chandra and Islam, Iftekharul and Wang, Xuan and Li, Weizi},
  journal={arXiv preprint arXiv:2504.05059},
  year={2025}
}