AUTODRAITEC: An Infrastructure-Based AUTOnomous DRiving System Using Artificial Intelligence and TEleCommunication Technologies

Zine El Abidine Kherroubi, Fouzi Boukhalfa, Thierry Lestable, Carlos-Faouzi BaderIJCAI 2024

Abstract

Highway on-ramp merging remains one of the most challenging maneuvers for autonomous vehicles (AVs) due to sensor limitations and the unpredictability of human drivers. This paper introduces AUTODRAITEC, an infrastructure-based system that shifts perception and decision-making from the vehicle to roadside edge computing. By utilizing a Bird’s-Eye-View (BEV) and a specialized Driver Intention Model (DIM), the system achieves a 30% improvement in safety distances and a 15% increase in merging efficiency compared to traditional human-driven scenarios.

Introduction

Despite rapid advancements, onboard sensors in autonomous vehicles often suffer from "tunnel vision"—limited range and blind spots caused by occlusions or environmental factors. These limitations are particularly dangerous at highway on-ramps, where merging vehicles must coordinate with high-speed traffic. Current AVs often struggle to predict the intentions of human drivers, leading to overly cautious or unsafe merging behavior.

AUTODRAITEC addresses these challenges by transforming the road into an intelligent infrastructure. By placing high-performance sensors and Multi-access Edge Computing (MEC) at the roadside, the system provides a comprehensive view of traffic, enabling centralized coordination that individual vehicles cannot achieve on their own.

The AUTODRAITEC Framework

The system architecture is built upon a three-stage pipeline deployed on roadside industrial computers:

  1. Perception and Data Fusion: The system integrates data from LiDAR and cameras. Using YOLOv5 for visual detection and CenterPoint/PointPillars for 3D point cloud processing, the data is fused via Bayesian statistics to maintain a 98% object detection accuracy.
  2. Driver Intention Model (DIM): To account for the "human factor," a supervised learning classifier predicts whether a human driver in the main lane intends to yield or maintain speed. This model achieved 99% accuracy, allowing the system to anticipate traffic gaps before they fully form.
  3. Actor Policy: A reinforcement learning agent, trained using the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, generates real-time speed recommendations for Connected and Autonomous Vehicles (CAVs) to execute smooth merges.

Experimental Methodology

To validate the system, the researchers developed a high-fidelity 1:18 scale physical testbed replicating a segment of Interstate 80 in Emeryville, California. This setup allows for the safe testing of complex traffic interactions using miniature Ackermann-steering vehicles.

The 1:18 scale physical testbed used to validate the AUTODRAITEC system, featuring highway and merge lanes equipped with roadside sensors and industrial computing.

The study also leveraged the NGSIM (Next Generation Simulation) dataset to ensure that the trajectories and traffic distributions in the scale model accurately reflected real-world highway conditions.

Results and Discussion

Evaluation across 10,000 test episodes demonstrated significant improvements in both safety and throughput. The system successfully eliminated collisions in scenarios where human drivers typically failed to coordinate safely.

  • Safety Enhancement: The average safety distance between vehicles increased by 30%.
  • Efficiency Gains: Merging speeds improved by 15%, reducing the congestion typically caused by hesitant merging.
  • Robustness: The DIM's 99% accuracy proved critical; by knowing a driver's intent, the system could proactively accelerate the CAV into a gap rather than waiting for the gap to be physically clear.
A comparison of merging outcomes: (i) human drivers frequently result in rear-end collisions in aggressive scenarios, while (ii) the AUTODRAITEC system coordinates safe merges by predicting driver intent.

Conclusion and Future Work

AUTODRAITEC demonstrates that infrastructure-based intelligence can significantly mitigate the risks of highway merging. By treating autonomous driving as a service provided by the road, the system enables safer and more efficient traffic flow. Future research will focus on scaling this approach to complex urban intersections and addressing the telecommunication latency requirements necessary for real-world 5G/6G deployment.

Summary updated

BibTeX
@inproceedings{kherroubi2024autodraitec,
  author = {Zine El Abidine Kherroubi and Fouzi Boukhalfa and Thierry Lestable and Carlos-Faouzi Bader},
  title = {{AUTODRAITEC: An Infrastructure-Based AUTOnomous DRiving System Using Artificial Intelligence and TEleCommunication Technologies}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure