This project builds a deep learning system that predicts how a driver is braking using vehicle time-series signals such as:
- Vehicle speed
- Acceleration (deceleration)
- Brake pedal input
The model predicts:
- Braking intention — Light / Normal / Emergency
- Brake intensity — a continuous value representing braking aggressiveness
Why this matters?
- Early braking prediction can improve vehicle safety systems and driver assistance technologies.
- If a car detects emergency braking earlier, safety systems can react faster.
- For electric vehicles, braking intensity prediction can also help optimize regenerative braking, improving energy recovery and efficiency.
| Model | Accuracy | Macro F1 | Normal Braking F1 | Emergency Braking F1 |
|---|---|---|---|---|
| Baseline (Single-task) | 69.6% | 70% | ~59% | ~78% |
| AE + Classifier (Best) | 64.1% | 64% | ~56% | ~77% |
| Multitask (λ = 0.5) | 69.0% | 70% | ~57% | ~77% |
| Multitask (λ = 0.8) | 71.3% | 72% | 59% | 82% |
Key result: Multitask learning improved overall performance and achieved 82% F1-score for Emergency Braking, the most safety-critical class.
Clone the repository and install dependencies:
git clone https://github.com/your-username/braking-intention-prediction.git
cd braking-intention-prediction
pip install -r requirements.txt(Optional) If using Jupyter notebooks:
python -m ipykernel install --user --name braking-intent- Generate Dataset
python data/generate_dataset.py
python data/generate_hard_dataset.pyThis will create .npy files containing time-series samples and labels.
- Train Baseline Model
Open and run:
01_train_baseline.ipynb
- Train Final Multitask Model
Open and run:
02_multitask_training.ipynb
All results, confusion matrices, and metrics are produced inside the notebooks.
- Run the Interactive Demo
streamlit run app.py-
Vehicle signals are synthetically generated to simulate realistic braking scenarios.
-
Short time windows of speed, acceleration, and brake input are sufficient to estimate braking intention.
-
Brake intensity correlates with braking aggressiveness.
- Model is trained on synthetic data; real-world deployment would require:
- Sensor calibration
- Domain adaptation
- Validation on real driving datasets
- Reaction latency, road conditions, and driver intent beyond braking are not modeled.
This project reproduces and extends ideas from the following research work:
Wei Yang, Yu Huang, Kongming Jiang, Zhen Zhang, Ketong Zong, Qin Ruan,
“Method of Predicting Braking Intention Using LSTM-CNN-Attention With Hyperparameters Optimized by Genetic Algorithm”,
International Journal of Control, Automation and Systems, Springer, 2024.
(https://link.springer.com/article/10.1007/s12555-021-1113-x)
The original paper proposes an LSTM–CNN–Attention architecture for braking intention prediction using simulator-based driving data.
This project reimplements the core architecture and introduces:
- harder ambiguous synthetic datasets
- systematic ablation studies
- multitask learning with braking intensity regression