How Does C-V2X Help Autonomous Driving to Avoid Accidents?

Lili Miao, Shang-Fu Chen, Yu-Ling Hsu, Kai-Lung Hua2022 · DOI

TL;DR

This paper proposes a Cellular Vehicle-to-Everything (C-V2X) based Vulnerable Road User Collision Warning (VRUCW) system to enhance autonomous driving safety, particularly in non-line-of-sight scenarios. The system leverages roadside units (RSUs) and onboard units (OBUs) to detect and warn vehicles about pedestrians and other road users, demonstrating its effectiveness in a real-world deployment.

What Problem Does This Paper Address?

Autonomous driving vehicles, while equipped with advanced sensors like cameras, radar, and LiDAR, still face significant challenges in ensuring safety, especially in adverse weather, poor lighting, or when encountering non-line-of-sight (NLOS) obstacles. These limitations can lead to accidents, as current sensor technology alone is insufficient to perceive all potential hazards. The paper addresses this gap by proposing a connectivity-based solution using C-V2X technology, which allows vehicles to communicate with infrastructure and other road users, thereby extending their situational awareness beyond the limitations of onboard sensors.

Methods

The core of the proposed solution is the Vulnerable Road User Collision Warning (VRUCW) system, which integrates C-V2X technology with autonomous driving systems. The system architecture involves:

  • Roadside Units (RSUs): These units, equipped with cameras and connected to an AI server, monitor the environment and detect vulnerable road users (VRUs). They broadcast this information to vehicles via the PC5 interface.
  • On-Board Units (OBUs): Installed in vehicles, OBUs receive information from RSUs and also gather data from the vehicle's control unit (GNSS and CAN messages). They then process this combined information to determine potential collision risks.
  • AI Server: This server processes video streams from IP cameras, employing AI algorithms (SSD and ResNet-18) for object detection, recognition, and movement prediction of VRUs.
  • Network Topology: A circular route is used for deployment, with RSUs strategically placed to cover intersections. IP cameras are connected via PoE switches to the AI server, forming the backbone of the detection system.
  • Message Flow: A defined message flow outlines how IP cameras send video to the AI server, which then relays detection results to the RSU. The RSU broadcasts this information to OBUs, which use this data along with vehicle telemetry to trigger warnings.
  • Security: A symmetric software encryption (AES-ECB) is implemented for C-V2X Sidelink transmission to ensure confidentiality and integrity, with keys manually shared between RSU and OBU.

Datasets

Not explicitly discussed in the paper. The AI models (SSD and ResNet-18) are trained on general object detection datasets, and the system's performance is evaluated in a real-world deployment scenario.

Key Insights and Ideas

  • C-V2X as a Complement to Onboard Sensors: The paper strongly advocates for C-V2X not as a replacement, but as a crucial enhancement to existing onboard sensors, particularly for addressing NLOS scenarios and improving safety in challenging conditions.
  • RSU-Based Detection for Enhanced Awareness: Leveraging RSUs equipped with cameras and AI processing allows for proactive detection of VRUs, even before they enter the vehicle's direct line of sight, significantly improving reaction times.
  • Practical Security for Early C-V2X Deployment: The proposed symmetric encryption offers a workable security solution for early-stage C-V2X deployments, addressing the lack of detailed security specifications in current V2X application layers.
  • Real-World Validation: The successful deployment and testing in a real village environment demonstrate the practical applicability and effectiveness of the proposed VRUCW system.

Key Results

  • Latency: The PC5 interface latency was tested with and without encryption. With encryption, the average latency was 20.53 ms, and without encryption, it was 20.83 ms, indicating that the added security function has a negligible impact on performance.
  • AI Classification Performance: The AI model achieved high precision and recall for detecting "Person" (98.61% precision, 95.73% recall) and "Vehicle" (91.79% precision, 99.70% recall), resulting in an average F1 score of 96.71%.
  • System Functionality: The VRUCW function was successfully demonstrated in a field experiment, where a "PERSON" warning message was displayed on the OBU's putty tool when a pedestrian was detected, and this warning was simultaneously sent to the autonomous driving control system.

Limitations and Future Work

  • Camera Coverage Limitations: The current system relies on a limited number of cameras, necessitating more cameras for full coverage, which increases investment costs. Future work suggests using smart cameras and LiDAR sensor fusion for more comprehensive detection.
  • Manual Geographic Information Configuration: The manual measurement and configuration of geographic data for the AI server can lead to deviations. Future work aims to automate this process using multiple probe tools.
  • End-to-End Latency: While PC5 latency is low, the overall end-to-end latency (including fiber-optic/cable transmission and AI computation) is within 200 ms. For more advanced services like remote driving, further latency reduction is needed.
  • Future Deployment: The authors plan to deploy AI workloads at the roadside smart sensor level to overcome latency issues and explore 5G NR-V2X for ultra-reliable low-latency communication.
  • Enhanced Security: While the current symmetric encryption is suitable for early deployments, future work will investigate software asymmetric encryption for more robust security.

Summary updated

BibTeX
@misc{miao2022how,
  author = {Lili Miao and Shang-Fu Chen and Yu-Ling Hsu and Kai-Lung Hua},
  title = {{How Does C-V2X Help Autonomous Driving to Avoid Accidents?}},
  year = {2022},
  doi = {10.3390/s22020686},
}
Figure