💻 computer science

A Hybrid Graph-Based and AI-Driven Cooperative System for Overtaking Collision Avoidance

This paper presents IG-COAS, a hybrid cooperative system that enhances overtaking safety on two-lane roads by integrating V2X communication, dynamic graph-based environment modeling, and cloud-assisted machine learning to provide real-time, predictive collision risk assessment and decision support.

Original authors: Safa Batita, Achraf Makni, Ikram Amous

Published 2026-08-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Safa Batita, Achraf Makni, Ikram Amous

Original paper licensed under CC BY 4.0 (https://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

Imagine you are driving down a two-lane road, the kind where the yellow line in the middle is the only thing keeping you from a head-on collision with a truck coming the other way. You want to pass a slow car ahead, but you can't see far enough around the bend to know if it's safe. This is the "overtaking problem," one of the most dangerous moves in driving. For decades, scientists have been trying to build "Advanced Driver Assistance Systems" (ADAS)—basically, super-smart co-pilots that can see what you can't. These systems rely on three big ideas: V2X communication (cars talking to each other like walkie-talkies), Graph Databases (a way of organizing data that looks more like a web of connections than a spreadsheet), and Artificial Intelligence (computers that learn from patterns to make predictions). The big question is: Can we combine these three tools into one super-system that doesn't just react to danger, but predicts it before it happens?

Enter the researchers Safa Batita, Achraf Makni, and Ikram Amous from the University of Sfax. They didn't just build a better co-pilot; they built a "digital nervous system" for the road called IG-COAS (Intelligent Graph model-based Cooperative Overtaking Assistant System). Think of it as a team of three: a local driver who reacts instantly, a cloud-based brain that learns from history, and a magical map that sees the invisible threads connecting every car.

Here's how their system works. Instead of just looking at a list of car speeds and positions, IG-COAS draws the traffic as a living, breathing graph. Imagine a spiderweb where every car, every lane, and every potential crash point is a knot, and the lines connecting them are the relationships between them. If Car A is behind Car B, there's a line. If Car C is coming from the opposite direction, there's a line that says "Danger Ahead." This web updates in real-time, creating a "Virtual Collision Point" (VCP)—a spot in the future where two cars might meet if they don't change their minds. It's like seeing a ghostly red X floating in the air where a crash would happen, allowing the system to warn you before you even get there.

To make this web smart, the system uses Artificial Intelligence. The researchers fed their AI a massive amount of simulated driving data, teaching it to spot the subtle signs of a risky overtake, like a car suddenly braking or speeding up too fast. They tested four different types of AI "brains" (KNN, Naive Bayes, Logistic Regression, and Support Vector Machines) to see which one was the best detective. The winner was a Support Vector Machine with an RBF kernel, which got it right 99.24% of the time in their tests. It was so good at spotting danger that it made very few mistakes, which is crucial because you don't want a system that screams "Crash!" when you're just changing lanes safely.

The team ran thousands of simulations on a two-lane road, cranking up the traffic from just 3 cars to a jam-packed 70 cars. They found that their system is incredibly fast. The "distributed" version (where the cars talk directly to each other without waiting for a central server) reacted in just 4.0 milliseconds with light traffic and only slowed down to 20.7 milliseconds even with 70 cars. Compare that to a "cloud-only" system, which took 39.0 milliseconds to start and dragged up to 77.0 milliseconds with heavy traffic. While 77 milliseconds is still faster than a human driver (who takes about 1,000 milliseconds to react), the direct car-to-car chat is clearly the speed champion for saving lives.

The results also showed a clear link between speed and danger. In their simulations, 50% of all the near-miss accidents happened when cars were traveling between 30 and 40 meters per second (about 67 to 90 mph). At lower speeds, the system rarely saw trouble. But as traffic got denser, the success rate of safe overtaking dropped from 85.1% in light traffic to just 34.3% in heavy jams, proving that while the system is helpful, physics still makes overtaking in a traffic jam a tough call.

So, what's the verdict? The paper suggests that IG-COAS is a promising new way to keep drivers safe. It successfully combines a real-time map of the road (the graph), a fast local reaction (the distributed system), and a smart learning brain (the AI) to spot overtaking risks earlier than current methods. However, the authors are careful to note that this is all based on simulations using tools like SUMO and OMNeT++. They haven't tested it on real roads with real cars yet. They also point out that while the cloud part is great for learning and storing data, it's too slow to be the main hero in a split-second emergency. The future, they suggest, lies in putting more computing power right on the cars (edge computing) and adding cameras and lasers to the mix. For now, IG-COAS stands as a strong, simulated proof that if we can get our cars to talk, map their connections, and learn from their mistakes, we might just be able to turn that dangerous yellow line into a safe path forward.

Technical Summary: A Hybrid Graph-Based and AI-Driven Cooperative System for Overtaking Collision Avoidance

Problem Statement
Overtaking on two-lane bidirectional roads remains a leading cause of fatal accidents due to limited visibility, dynamic vehicle interactions, and the difficulty of accurately estimating collision risk. Existing Advanced Driver Assistance Systems (ADAS) often rely solely on onboard sensors, which are limited by occlusion and environmental conditions, or they utilize isolated communication or data-driven components without a unified architecture. Current approaches frequently lack the ability to merge real-time data consistency, graph-based modeling of complex vehicle relationships, cooperative communication, and cloud-assisted intelligence into a single framework capable of predicting and mitigating overtaking risks effectively.

Methodology
The paper proposes the Intelligent Graph model-based Cooperative Overtaking Assistant System (IG-COAS), a hybrid framework designed to integrate distributed vehicle-level responsiveness with cloud-based learning. The methodology is structured around four core components:

Hybrid Architecture: The system operates on two tightly linked layers:
- Vehicular Layer: Acts as the real-time operational core, gathering motion states (position, speed, direction) and maneuver intentions (e.g., turn signals). It utilizes a Real-Time Graph Database (RT-GDB) to maintain a dynamic representation of the traffic environment.
- Cloud Layer: Handles large-scale data storage, model training, and long-term behavioral analysis. It continuously refines predictive models based on aggregated vehicle data, which are then deployed back to the vehicle edge.
Graph-Based Environment Modeling: The traffic environment is modeled as a directed graph $G = (V, E)$ .
- Nodes: Represent vehicles, lanes, and Virtual Collision Points (VCP). VCPs are dynamically generated to predict potential frontal conflicts between overtaking and oncoming vehicles.
- Edges: Encode semantic and temporal relationships, such as LEADER (longitudinal ordering), OPPOSITE DIRECTION (conflicting flows), and NEXT (arrival sequencing at a VCP).
- Temporal Validity: To ensure data freshness, attributes are treated as temporal quadruplets $(d_{value}, d_{timestamp}, d_{duration}, m_{de})$ , allowing the system to assess the validity of information and trigger selective updates only when values deviate significantly or validity periods expire.
Machine Learning Risk Assessment: A supervised learning module classifies overtaking scenarios as "safe" or "unsafe."
- Data Generation: Due to the lack of public datasets capturing specific overtaking behaviors, data was generated using a simulation framework combining OpenStreetMap (OSM) for road geometry and SUMO for traffic simulation, integrated with OMNeT++ and VEINS for V2X communication.
- Feature Engineering: The system derives critical safety indicators not directly available in standard Floating Car Data (FCD), including deceleration, relative speed, Time-To-Collision (TTC), and safe distance.
- Model Selection: Four algorithms (KNN, Naive Bayes, Logistic Regression, SVM) were evaluated. The Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel was selected as the optimal model, achieving 99.24% accuracy, 98.92% precision, and 99.73% recall.
Decision Logic: The system employs a hybrid decision process (Algorithm 1) that first checks for overtaking intent. If an intent is detected, it verifies kinematic safety conditions (e.g., TTC < 3s, LeaderGap < SafeDistance). If risks are flagged, the ML module provides a secondary prediction to confirm the decision before triggering a warning or responsive action.

Key Contributions
The paper claims the following contributions:

Unified Hybrid Architecture: A cohesive system integrating cloud-level learning with distributed vehicular agents for real-time risk assessment, addressing the gap between centralized intelligence and distributed autonomy.
Dynamic Graph Representation: A novel modeling approach where vehicles, lanes, and collision points are interconnected entities enriched with temporal metadata, enabling the system to view overtaking not as isolated actions but as evolving interaction patterns.
Cloud-Edge ML Integration: A risk classifier trained in the cloud and implemented at the vehicle edge, bridging high-accuracy predictive models with real-time execution constraints.
Cooperative Decision Support: A process leveraging V2X communication and shared traffic-state information to enhance situational awareness beyond local sensor capabilities.

Experimental Results
The system was evaluated through extensive simulations under varying traffic densities (3 to 70 vehicles) and speeds (up to 40 m/s). Key findings include:

Reaction Time (T-react): The distributed architecture demonstrated the lowest reaction times, ranging from 4.0 ms (low density) to 20.7 ms (high density). The ML-assisted configuration ranged from 8.0 ms to 35.9 ms, while the cloud-only configuration was significantly slower (39.0 ms to 77.0 ms), confirming its suitability for non-time-critical tasks like data storage rather than immediate decision-making.
Accuracy: The SVM-RBF model achieved a 99.24% accuracy in classifying overtaking safety, with only 12 false negatives and 49 false positives in the test set.
Traffic Density Impact: As traffic density increased, the average success rate of overtaking maneuvers decreased from 85.1% (low density) to 34.3% (high density), highlighting the system's ability to identify increasingly constrained and risky environments.
Speed Correlation: The analysis revealed a strong correlation between vehicle speed and accident frequency, with 50% of detected overtaking incidents occurring in the 30–40 m/s range.

Significance
The paper positions IG-COAS as a significant step toward next-generation ADAS by demonstrating that integrating graph-based modeling, cooperative communication, and AI-driven prediction into a unified framework can effectively identify unsafe overtaking situations earlier than existing approaches. The authors claim the system improves reaction times and maintains robustness across varying traffic densities. While the results are promising, the paper maintains a modest tone regarding real-world deployment, noting that current limitations include reliance on simulation and the need for future work to integrate multimodal sensors (cameras, LiDAR) and edge computing to further reduce latency and enhance situational awareness. The system is presented as a foundational step toward safer, smarter roads, pending further development and real-world validation.

More like this