Enhancing Traffic Safety Through V2X-Based Deep Reinforcement Learning

Abstract

As a core component of an ITS, CAVs have garnered significant attention and expect to play an important role in enhancing traffic safety. To enable (future) practical deployment of CAVs, safety strategies must be incorporated into the control design. This includes, in particular, strategies for ensuring avoidance or mitigation of multi-vehicle collisions. Conventional control strategies exhibit some inherent limitations that limit their ability to improve efficiency and safety beyond a certain point. This thesis proposes DRL as strategy or means to go beyond the limited performance of classic control strategies. Our preliminary results demonstrate the potential of using DRL for design of emergency braking profiles in vehicle-following scenarios. In the studied scenarios, three vehicles are involved, where the middle vehicle shall decelerate in such a way that collective harm is minimized for the three vehicles involved. Based on the developed DRL-approach, this thesis further provides a hybrid approach that combines DRL with an existing method based on analytical expressions for selecting optimal constant deceleration for the middle vehicle. By combining DRL with the previous method, the proposed hybrid approach increases the reliability compared to standalone DRL, while achieving superior performance in comparison to the optimal constant deceleration approach in terms of overall harm reduction and collision avoidance.

Popular Abstract

Self-driving and connected cars are expected to play a key role in making our roads safer. But for them to be used widely in the future, they must be equipped with smart ways to control them to prevent or reduce the impact of accidents—especially when several vehicles are involved. Traditional control methods, while useful, have limits on how much they can improve safety and efficiency. This research explores using deep reinforcement learning (DRL), a branch of artificial intelligence that learns strategies through trial and error. Our study focuses on emergency braking situations in a three-car scenario, where the vehicles drive after each other on a road. Imagine that the first car suddenly needs to emergency brake, for example due to a suddenly appearing pedestrian on the road. This decision affects both the car in the middle and the last car. The challenge we are looking at is how the middle car should brake in controlled fashion so that the total harm (collision damage) for all three vehicles is minimized. We developed a DRL-based approach to handle this problem and found that it can produce more effective braking strategies than traditional methods. To further improve reliability, we combined DRL with an existing approach that calculates optimal braking under constant deceleration. This hybrid approach not only avoids collisions more effectively but also reduces overall harm better than either method on its own. In short, our work shows that combining AI with proven more classic method could pave the way for safer and more efficient driving.

• author: Wang, Jianbo
• supervisor: Johan Thunberg; Galina Sidorenko
• organization: Department of Electrical and Information Technology
• course: EITM02 20251
• year: 2025
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• keywords: Multi-vehicle safety, reinforcement learning, vehicle-to-vehicle communication, collision avoidance, automated driving.
• report number: LU/LTH-EIT 2025-1097
• language: English
• id: 9213360
• date added to LUP: 2025-10-17 09:47:59
• date last changed: 2025-10-17 09:47:59