Traffic Forecasting using Deep Sequence Models with Vehicle Situation-aware Loss (VSAL)

Official Implementation of "Traffic Forecasting using Deep Sequence Models with Vehicle Situation-aware Loss"
Akash Chatterjee, Jayant Mahawar, and Angshuman Paul
Indian Institute of Technology Jodhpur, India

---

📋 Table of Contents

• Overview
• Key Features
• Architecture
• Installation
• Dataset
• Usage
• Results
• Citation
• Acknowledgments
• License

---

🔍 Overview

This repository presents a novel approach to traffic forecasting that addresses the limitations of existing methods by introducing the Vehicle Situation-aware Loss (VSAL)—a composite loss function that enables deep sequence models to simultaneously learn multiple interdependent traffic variables.

The Problem

Traditional traffic forecasting models often:

• Focus solely on velocity prediction
• Ignore crucial factors like lane changes and traffic density
• Rely on spurious correlations with limited behavioral understanding

Our Solution

VSAL integrates multiple loss components to holistically capture:

• ✅ Vehicle velocity and position
• ✅ Geospatial accuracy (latitude/longitude)
• ✅ Lane-change classification
• ✅ Traffic congestion estimation
• ✅ Self-consistency between predicted velocity and displacement

---

⭐ Key Features

• Novel Composite Loss Function: VSAL combines six distinct loss components for comprehensive traffic modeling
• Enhanced Model Variants: VS-LSTM, VS-GRU, and VS-Transformer architectures
• Rich Contextual Features: Lane-specific gap distances, traffic density, speed reduction, and novel Jam Factor
• State-of-the-art Performance:
  • VS-Transformer achieves 0.002 velocity RMSE (99.9% improvement over baseline)
  • 0.358 Haversine RMSE for geospatial accuracy
  • 77.2% lane-change classification accuracy
• Multiple Prediction Tasks: Simultaneous prediction of velocity, position, coordinates, and lane changes

---

🏗️ Architecture

VSAL Components

The Vehicle Situation-aware Loss consists of:

L = L_vel + α·L_pos + β·L_sc + γ·L_class + δ·L_cong + σ·L_hav

1. Velocity Loss (L_vel): MSE for speed prediction
2. Position Loss (L_pos): MSE for longitudinal position
3. Self-Consistency Loss (L_sc): Ensures predicted position matches velocity-based displacement
4. Lane-Change Classification Loss (L_class): Cross-entropy for lane-change detection
5. Congestion Prediction Loss (L_cong): MSE for Jam Factor prediction
6. Haversine Loss (L_hav): Geospatial distance for coordinate accuracy

Model Architectures

Input Features → Sequence Model → Multi-Task Heads → Predictions
     ↓                ↓                    ↓              ↓
  velocity      LSTM/GRU/        Regression Head    velocity
  position      Transformer      Classification     position
  gaps                           Congestion Head    lat/long
  density                                          lane change
  Jam Factor                                       congestion

---

🛠️ Installation

Requirements

• Python 3.8+
• PyTorch 2.0+
• CUDA 11.0+ (for GPU support)

Setup

# Clone the repository
git clone https://github.com/yourusername/traffic-forecasting-vsal.git
cd traffic-forecasting-vsal

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Dependencies

torch>=2.0.0
numpy>=1.21.0
pandas>=1.3.0
matplotlib>=3.4.0
scikit-learn>=1.0.0
pyproj>=3.6.1
tqdm>=4.62.0

---

📊 Dataset

This project uses the NGSIM US-101 trajectory dataset, which contains high-fidelity vehicle trajectory data collected on the southbound US 101 freeway in Los Angeles.

Dataset Preparation

# Download NGSIM US-101 dataset
python scripts/download_dataset.py

# Preprocess the data
python scripts/preprocess_data.py

Data Preprocessing Pipeline

1. Data Cleaning: Remove erroneous lanes (6, 7, 8)
2. Unit Conversion: Convert feet to meters
3. Temporal Sorting: Sort by timestamp
4. Feature Extraction:
   • Lane-specific gap distances (g1-g6)
   • Traffic density per lane
   • Speed reduction metric
   • Jam Factor computation
5. Coordinate Transformation: NAD83 (EPSG:2227) → WGS84 (EPSG:4326)
6. Normalization: Min-max scaling

---

🚀 Usage

Training

# Train VS-LSTM
python train.py --model lstm --use_vsal --epochs 50 --batch_size 64 --lr 0.0005

# Train VS-GRU
python train.py --model gru --use_vsal --epochs 50 --batch_size 64 --lr 0.0005

# Train VS-Transformer
python train.py --model transformer --use_vsal --epochs 50 --batch_size 64 --lr 0.0005

Training with Custom Loss Weights

python train.py --model transformer \
                --use_vsal \
                --alpha 1.0 \
                --beta 1.0 \
                --gamma 1.0 \
                --delta 1.0 \
                --sigma 1.0

Evaluation

# Evaluate trained model
python evaluate.py --model_path checkpoints/vs_transformer_best.pth \
                   --data_path data/test.csv

Inference

# Run inference on new data
python inference.py --model_path checkpoints/vs_transformer_best.pth \
                    --input_data sample_trajectory.csv \
                    --output_dir predictions/

---

📈 Results

Quantitative Performance

Model	Velocity RMSE ↓	Local_Y RMSE ↓	Haversine RMSE ↓	Congestion RMSE ↓	Lane Change Acc ↑
Baseline LSTM	3.139±2.518	177.641±51.516	175.935±51.698	0.295±0.038	0.468±0.073
VS-LSTM	0.050±0.042	0.277±0.005	0.240±0.000	0.219±0.036	0.700±0.063
Baseline GRU	3.231±2.052	170.273±12.954	161.020±7.397	0.069±0.022	0.491±0.007
VS-GRU	0.004±0.002	0.264±0.007	0.240±0.000	0.068±0.002	0.772±0.001
Baseline Transformer	10.480±0.009	179.313±0.029	54.511±0.008	0.024±0.000	0.366±0.002
VS-Transformer	0.002±0.001	0.297±0.008	0.358±0.111	0.005±0.002	0.772±0.001

Key Improvements

• 🎯 98.4-99.9% reduction in velocity RMSE across all models
• 🎯 99.8% reduction in positional error
• 🎯 77.2% lane-change accuracy for VS-GRU and VS-Transformer
• 🎯 98% reduction in congestion prediction error

Visualization

Click to view sample predictions

Predicted vs Actual Vehicle Velocity showing close alignment even during rapid fluctuations

Longitudinal position prediction demonstrating spatial accuracy

Jam Factor prediction capturing congestion onset and dissipation

---

🔬 Ablation Studies

Each component of VSAL contributes uniquely to model performance:

Removed Component	Impact
Without L_pos	Decreased positional accuracy
Without L_sc	Increased velocity inconsistency
Without L_class	Significant drop in lane-change detection
Without L_cong	Poor congestion estimation
Without L_hav	Reduced geospatial precision

---

📝 Citation

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

@inproceedings{chatterjee2025traffic,
  title={Traffic Forecasting using Deep Sequence Models with Vehicle Situation-aware Loss},
  author={Akash Chatterjee, Jayant Mahawar and Angshuman Paul},
  booktitle={CVIP 2025},
  year={2025},
  organization={Indian Institute of Technology Jodhpur}
}

---

🤝 Acknowledgments

• NGSIM dataset provided by the Federal Highway Administration
• Research conducted at Indian Institute of Technology Jodhpur
• Thanks to the open-source community for PyTorch and related tools

---

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

👥 Authors

• Akash Chatterjee - GitHub | Email
• Jayant Mahawar - GitHub | Email

---

📧 Contact

For questions or collaboration opportunities, please contact:

• Email: [email protected]
• Institution: Indian Institute of Technology Jodhpur

---

⭐ Star History

If you find this project useful, please consider giving it a star! ⭐

---

Made with ❤️ at IIT Jodhpur