DRIFT: Driving Risk Inference via Field Transmission for Human-like Autonomous Driving

This repository is the official implementation of the paper "DRIFT: Driving Risk Inference via Field Transmission for Human-like Autonomous Driving", accepted for publication at The IEEE International Conference on Intelligent Transportation Systems (ITSC) 2026

Key Features

• Unified PDE Framework: Single equation captures advection, diffusion, and wave-like propagation
• GVF-based Vehicle Interaction: Anisotropic Gaussian kernels model relative motion risk
• Occlusion Reasoning: Shadow regions behind trucks inject latent hazard
• Merge Topology: Road geometry creates conflict zones with elevated risk
• Interpretable Sources: Clear decomposition into Q_veh, Q_occ, Q_merge

Source Terms

Source	Description
Q_veh	Vehicle-induced risk using GVF-style anisotropic Gaussians weighted by TTC, relative speed, and vehicle class
Q_occ	Occlusion hazard in sensor shadow behind large vehicles; higher at lane centers and truck edges where cut-ins emerge
Q_merge	Merge conflict pressure; intensifies toward gore point with topology-driven drift

Dataset loading and Sceanario Engineering

see another repo: https://github.com/PeterWANGHK/Benchmark-RiskField.git

Getting started

1. Install dependencies:

   git clone https://github.com/PeterWANGHK/DRIFT.git
   pip install numpy scipy matplotlib imageio loguru

2. Run verification:

   cd DRIFT/src
   python test_pde_fixes.py

3. Generate visualization:

   python drift_pde_visualization.py

4. Loading sceanrioos from BEV dataset if needed (please specify the dataset directory in corresponding code lines in your drif_pde_xxd.py)

   #example usage of exiD dataset
   python drift_pde_exid.py --recording 00 --ego_id 5

5. Fine-tune parameters if needed (see tuning guide in another branch)

Demonstration example:

single vehicle field propagation (with truck occlusion effects)

single vehicle with static occluder (obstacle)

Group field propagation:

Group field propagation with occlusion-aware and merging pressure:

Ablation study: the effects of each contributing term (advection, diffusion)

python drift_pde_visualization.py --ablation --frames 70 --fps 8

evaluation in different scenarios using BEV datasets: