Scalable Edge-assisted Fusion and Path Prediction for Connected Autonomous Vehicles
This paper presents Conductor, an edge-assisted system for Connected Autonomous Vehicles that dynamically selects occlusion-aware vehicle inputs and adapts fusion workloads to maintain strict Age of Information bounds while generating a unified world model and trajectory predictions that significantly outperform state-of-the-art local fusion methods.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Self-driving cars rely on a constant stream of sensory data to navigate safely, but their cameras and lasers have a fundamental weakness: they cannot see through walls, parked trucks, or other vehicles. When a pedestrian steps out from behind a bus, or a car speeds around a blind corner, the autonomous vehicle's own sensors often do not notice the danger until it is too late to stop. To solve this, engineers are building networks where cars and roadside units share what they see, creating a collective view of the road that is far more complete than any single vehicle could achieve alone. However, this shared information is only useful if it arrives instantly. If the data is even a fraction of a second too old, the car's computer might be making decisions based on a world that has already changed, turning a helpful alert into a dangerous distraction.
Researchers at the Georgia Institute of Technology and the University of Edinburgh have developed a new system called Conductor to manage this delicate balance. They discovered that simply adding more cars to the network does not always make it safer; in fact, trying to process data from too many vehicles at once can slow down the system so much that the information becomes stale before it reaches the driver. In their simulations, when more than seventeen cars tried to share their views simultaneously, the time it took to process and deliver the data crossed a critical safety threshold of three hundred milliseconds. At typical city speeds, this delay represents a distance of four meters, or roughly one lane width, which is the exact margin needed to brake safely in an emergency. The researchers found that this "deadline cliff" happens because the computer at the roadside gets overwhelmed trying to merge all the incoming streams of information at once.
To prevent this, the team designed a smart filter that acts as a traffic controller for the data. Instead of accepting information from every available car, the system analyzes which vehicles are actually seeing something the roadside camera cannot. If a car is parked in a spot where its view duplicates what the roadside unit already sees, the system ignores it. It only selects the cars that are spotting hidden objects, such as a vehicle obscured by a building or a pedestrian in a blind spot. This selection process happens in real-time, ensuring that the system only processes the most valuable new information. Once the most useful cars are chosen, the system also adjusts how much it predicts the future movements of other vehicles. If the processing time is running tight, it focuses its computing power on the most dangerous or unpredictable objects, while letting simpler, faster calculations handle the rest.
The researchers tested this approach in a simulated environment that mimicked busy city intersections with up to thirty-one connected cars. They compared their new method against older approaches that either tried to use every single car or relied on simpler, less detailed data sharing. The results showed that their system was the only one capable of keeping the information fresh enough to be safe across all traffic densities. In scenarios where the roadside camera had large blind spots, the system's ability to pick the right cars to watch allowed it to recover eighty-seven percent of the hidden objects that a random selection of cars would have missed. When they tested the system in a closed-loop driving simulation, where the virtual cars had to actually drive and react to the data, the new method allowed the vehicles to navigate complex, crowded intersections without crashing or making sudden, unnecessary stops. In contrast, systems that tried to use all available data failed under heavy traffic loads, and systems that relied on simple data sharing failed to see hidden dangers.
The study confirms that for connected autonomous vehicles to be truly safe, the technology must be as much about managing time and computing resources as it is about seeing the road. By carefully selecting which cars contribute to the shared view and adapting the complexity of the predictions to the available time, the Conductor system ensures that the collective intelligence of the network remains a reliable partner for the driver, rather than a source of dangerous delays. This work suggests that the future of autonomous driving lies not just in building smarter sensors, but in building smarter ways to share the data those sensors collect.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.